Research on the Construction and Application of a SVM-Based Quantification Model for Streetscape Visual Complexity

Abstract

Visual complexity is a crucial criterion for evaluating the quality of urban environments and a key dimension in arousal theory and visual preference theory. Objectively quantifying visual complexity holds significant importance for decision-making support in urban planning. This study proposes a visual complexity quantification model based on a support vector machine (SVM), incorporating six key indicators, to establish a mapping relationship between objective image features and subjective complexity perception. This model can efficiently and scientifically predict street view complexity on a large scale. The research findings include the following: (1) the introduction of a new quantification dimension for the urban environment complexity—hierarchical complexity– which reflects the richness of street elements based on an in-depth semantic understanding of images; (2) the established complexity quantification model demonstrates high accuracy, with the indicators ranked by contribution for compression ratio, grayscale contrast, hierarchical complexity, fractal dimension, color complexity, and symmetry; and (3) the model was applied to predict and analyze the visual complexity of the Xiaobailou and Wudadao Districts in Tianjin, revealing that the visual complexity of most streets is moderate, and targeted recommendations were proposed based on different levels of visual complexity.

1. Introduction

As cities continue to develop, a notable shift from urban expansion to urban renewal has occurred. The increasing demand for urban environmental quality and livability has become more pronounced [1,2]. In this new urban development phase, visual-spatial quality and the built environment have become top priorities [3,4]. Visual complexity, a key concept in environment perception, is pivotal in building design and assessing environment visual quality. It not only shapes people’s emotional responses and aesthetic experiences [5] but is also closely linked to the visual appeal of the environment [6], its therapeutic benefits [7], and individual behavior patterns [8]. Thus, scientifically measuring visual complexity is crucial for creating high-quality urban built environments.

Although methods for quantifying visual complexity have been widely applied in fields such as cognitive psychology [6], computer science [5,9], and the arts [10,11], research in environment visual studies remains limited, and a comprehensive methodological system has yet to be established. Most studies still rely on manual evaluation as the primary data collection method. There are two main approaches: (1) expert-based objective evaluation, which follows predefined standards to systematically assess visual complexity based on factors such as image color and landscape elements; and (2) public-based subjective evaluation [12,13], which often encounters challenges such as high time consumption, limited evaluation scope, delayed feedback, and interference from subjective preferences.

As machine learning and computer vision technologies continue to advance, methods for evaluating image complexity are also evolving. Many studies have begun to employ machine learning techniques to assist manual scoring [14,15], aiming to evaluate large-scale datasets efficiently. However, this approach often faces challenges in terms of accuracy. Meanwhile, assessment methods based on the combination of image features have emerged as the mainstream [16,17], as these models are better at understanding and learning the underlying complex patterns in the data. Additionally, the decision-making process of such models is more transparent, and they tend to achieve higher accuracy. Cavalcante et al. utilized spatial frequency statistics and local contrast to calculate perceived complexity in the streetscape, but their method has limitations regarding stability and accuracy across different street types [18]. Liang et al. integrated traditional assessment methods with machine learning techniques to analyze six perceptual attributes of streetscape images, demonstrating that images with higher levels of complexity tend to elicit more positive perceptions [19]. Conversely, Ma et al. classified streetscape types based on three-dimensional features of complexity—texture, shape, and color—and developed a digital streetscape indexing system. However, their primary focus was not on quantifying visual complexity [20]. Supervised learning, a branch of artificial intelligence, has demonstrated substantial potential in simulating human behavior and decision-making, especially in integrating multiple features and addressing complex problems.

Visual complexity is characterized by a high density of available information, where the ratio of available to total information in an environment, when optimized, enhances its attractiveness and arousal potential for individuals [21]. Psychologist Mehrabian proposed that environmental information primarily stimulates individual responses through three dimensions: intensity, complexity, and novelty [22]. As a critical indicator influencing visual perception, defining complexity is a fundamental prerequisite and a key focus for conducting quantitative research. Kaplan and Kaplan (1989) linked complexity with the number and organization of visual elements in an environment, suggesting that highly complex yet coherent landscapes can provide visual richness, thereby enhancing visual aesthetics [23,24]. Ode, Tveit, and others defined complexity as the diversity and richness of landscape elements, as well as the dispersion of visual patterns and variability of features [25]. Ewing viewed complexity as visual richness, emphasizing the importance of diversity in the physical environment [26]. Most experts consider complexity from the perspective of human subjective experience, focusing on its influence on environment perception and aesthetic experience. They connect complexity with visual characteristics and environmental perception, highlighting the contribution of diversity and richness to visual complexity [15,27].

Additionally, visual perception is driven by two strategies [28]: bottom-up, which is initiated by low-level image features such as color and texture, and top-down, which is guided by high-level semantic information [29,30]. Traditional methods for quantifying subjective complexity rely primarily on low-level visual features, which present certain limitations. To more accurately reflect human high-level perception, this study introduces hierarchical complexity based on semantic image segmentation as a form of high-level semantic information. This innovative approach simulates human understanding of images by extracting multi-level semantic information, thus bridging the gap between low-level visual features and high-level human semantics [31].

As illustrated in Figure 1, this study introduces a novel method for measuring visual complexity. It employs machine learning to establish a mapping between objective image features and subjective complexity perception, enabling the prediction of subjective streetscape complexity based on objective image characteristics. This method efficiently processes large-scale datasets. The study identifies six key indicators—compression ratio, symmetry, fractal complexity, color complexity, grayscale contrast, and hierarchical complexity—and utilizes multi-feature fusion training. This approach offers a comprehensive explanation of the multidimensional nature of image complexity, closely mirroring human visual perception. The model was subsequently applied to assess the visual complexity of streetscapes in key districts of Tianjin. The primary contributions of this research are as follows:

• Introduction of a new measurement dimension: The study proposes hierarchical complexity as an advanced image feature to more accurately capture the intricacies of street scenes while bridging semantic gaps and validating its effectiveness.
• Development of a quantitative model: A quantitative model for streetscape visual complexity is constructed, with an analysis of the contributions of various indicators.
• Empirical validation and geographic analysis: The quantitative model is empirically validated, and an in-depth analysis of the geographic distribution of visual complexity is conducted within Tianjin’s Xiaobailou and Wudadao Districts.

2. Materials and Methods

2.1. Data Acquisition

With the rapid advancement of computer technology and the emergence of new techniques, the methods for acquiring streetscape images and processing visual information have expanded significantly. Using map APIs to obtain streetscape data for landscape analysis has become a mainstream approach, with proven accuracy and reliability in detecting urban environmental changes [32,33]. This study focuses on the Xiaobailou and Wudadao Districts in Tianjin’s Heping District (Figure 2). These areas, which are among Tianjin’s five characteristic districts, are representative of the city’s urban landscape. They feature a wealth of exquisite historical buildings with diverse architectural styles and varied street landscapes, making them ideal for analyzing visual complexity.

In streetscapes, a human-centered perspective is not only the most directly observable viewpoint in daily life, closely related to human well-being and health, but also serves as a critical complement to in-depth analysis of urban planning and design [34,35]. In this study, streetscape images were acquired using the Baidu Maps API, sampling at 40-m intervals. For each sampling point, images were captured from four angles—0°, 90°, 180°, and 270°—to provide a comprehensive view (Figure 3). The most recent summer street view images (2015–2020) were selected to minimize the influence of seasonal and other factors on the results.

A total of 1811 sampling points were collected across the Xiaobailou and Wudadao Districts, yielding 7244 images. After excluding images that were overexposed, underexposed, or heavily obstructed, 6729 high-quality streetscape images were retained. To ensure the model could learn diverse street scene features, 560 images were randomly selected for the training dataset, while the remaining 6169 images were used for further predictions, enhancing the model’s generalization ability in practical applications.

2.2. Identification of Complexity Quantification Features

2.2.1. Compression Ratio

In a 2005 study, Donderi and McFadden highlighted a significant correlation between image compression methods, such as JPEG and ZIP, and subjective assessments of image complexity [36]. The compression method reduces data by identifying repetitive patterns and statistical features within an image. However, since the initial file sizes of different images vary, directly comparing their complexity using this method may lack objectivity. The compression ratio—the ratio of compressed data size to the original data size—emerged as an effective metric for quantifying image complexity. Several studies have also demonstrated that the compression ratio can reflect the complexity of an image to some extent, with different images exhibiting varying compression ratios [37,38].

Street view images obtained through the Baidu Maps API typically undergo standardized processing to ensure that the same compression and optimization methods are applied to all image requests. This ensures consistency and comparability of streetscape images from different locations. In this study, we used the JPEG format for the compression, which can identify and remove redundant information within an image, thereby reducing file size while preserving visual quality as much as possible. Simple images are easier to compress because they typically contain a significant amount of redundant and predictable information, which compression algorithms can recognize and exploit to reduce data size effectively. Furthermore, to establish a positive correlation between the compression ratio and perceived complexity, we defined the compression ratio as the size of the compressed image relative to the original image size. Thus, the more complex the image, the lower the compression ratio.

2.2.2. Symmetry

Symmetry is a critical factor influencing aesthetic appeal. Numerous studies have demonstrated that the perception of complexity is closely related to elements of order and asymmetry within a scene [11,39], which, in turn, is strongly linked to visual environment preferences and widely applied in visual arts and spatial design [40,41]. Symmetry is closely associated with consistency in images, which is intrinsically connected to complexity [23]. Bigoin-Gagnan and Lacoste-Badie, in their study on packaging design, showed that complexity is affected by symmetry [40], while Chen et al. confirmed that symmetry significantly influences perceived visual complexity in natural images [42].

Streetscape images often contain complex backgrounds and various elements such as buildings, roads, and trees. Traditional methods for quantifying symmetry, such as Fourier transform, Hough transform, feature point-based methods, and convolutional neural networks, have shown limited effectiveness in detecting symmetry in street scenes. These methods suffer from high computational costs and are susceptible to interference from background and noise.

This study evaluates image symmetry by calculating the similarity of grayscale histograms. Considering that streetscape images typically exhibit horizontal distribution, we focus on comparing the similarity between the left and right sides of the image.

Specifically, the image width is denoted as $W$ Starting from the ${\displaystyle \frac{W}{4}}$ position along the width of the image, we iteratively adjust the cutting position until reaching ${\displaystyle \frac{3W}{4}}$, ensuring that the sizes of the left and right halves are approximately equal. The similarity between the grayscale histograms of the left and right sections is then calculated using the Bhattacharyya distance, $d_B\left(H_L\left(s\right),H_R\left(s\right)\right)$. The symmetry score is defined as the complement of the Bhattacharyya distance, as follows:

$$\mathrm{Symmetry} \mathrm{Score} \left(\mathrm{s}\right)=1-d_B\left(H_L\left(s\right),H_R\left(s\right)\right)$$

(1)

The larger the complement, the more similar the two probability distributions are. Finally, by iteratively adjusting the cutting position ‘s’, the highest symmetry score and the optimal cutting position are updated. The highest symmetry score is then returned as the final symmetry result for the image. The calculation formula is as follows:

$$\mathrm{Max} \mathrm{Symmetry} \mathrm{Score}=\underset{s\in \left[\frac{W}{4},\frac{3W}{4}\right]}{max}\left(1-d_B\left(H_L\left(s\right),H_R\left(s\right)\right)\right)$$

(2)

This method is computationally efficient and suitable for processing large-scale image data, and both grayscale histograms and Bhattacharyya distance are robust to noise and small local variations, effectively handling complex street scene images. The results are shown in Figure 4.

2.2.3. Fractal Dimension

Fractal dimension is a critical metric for quantifying the complexity of image content. It effectively reflects visual complexity in natural and built environments [43]. Fractal complexity is a significant factor influencing visual restorative properties in both natural and architectural settings [44]. Rich architectural ornamentation can achieve a similar level of attractiveness and restorative benefits as natural environments [45]. The fractal dimension focuses on the edges and structural information of image elements, making it suitable for describing the complexity of irregular shapes.

The box-counting method is a commonly used technique for calculating fractal dimensions [43,46]. This method involves covering a streetscape image $O$, which has been processed using the Canny edge detection method, with squares of varying side lengths (${\epsilon }_1,{\epsilon }_2,{\epsilon }_3,{\epsilon }_4,\cdots$) The number of squares required to cover the image at each size ($N\left({\epsilon }_1\right),N\left({\epsilon }_2\right),N\left({\epsilon }_3\right),N\left({\epsilon }_4\right),\dots$) is recorded. Typically, a linear relationship is observed between $\log N\left(\epsilon \right)$ and $\log \left(1/\epsilon \right)$, with the slope of this linear relationship representing the fractal dimension of the image $O$. The formula for calculating the fractal dimension is as follows:

$$D\left(O\right)=\underset{\epsilon \to \infty }{lim}{\displaystyle \frac{log N\left(\epsilon \right)}{log\left(\frac{1}{\epsilon }\right)}}$$

(3)

2.2.4. Color Complexity

Tsutsui and Ohmi developed a scale for evaluating the complexity of color combinations, demonstrating that the subjective complexity of color combinations can be defined by the relationships between colors (contrast or similarity) and the number of colors [47]. Color complexity focuses on the distribution and variation of colors within an image, and images’ complexity can be assessed by analyzing pixel color values [48]. Building on the findings of Zhou et al., this paper highlights that both color diversity and the spatial distribution of colors significantly contribute to color complexity [49]. The color complexity $C$ is more accurately assessed by considering the combined effects of color diversity $C_s$ and spatial distribution complexity $C_d$:

$$C=C_s+C_d$$

(4)

In contrast to the RGB color model, the HSV model exhibits a stronger intuitive correspondence between hue (H) and human color perception, where S represents saturation, and V represents value (brightness) [50]. In this study, each color $C_i$ is defined by quantizing the hue channel. The images are loaded using the OpenCV library in Python 3.10.0, and the RGB color space is converted to the HSV color space. The number of $bins$ represents the quantization intervals, and after quantizing the hue, each color $C_i$ is represented by the following formula:

$$C_i={\displaystyle \frac{H·bins}{180}}$$

(5)

where $N$ is the total number of pixels, and $n_i$ is the number of pixels of the $i^{th}$ color. There are $k$ distinct colors in total. The quantification of color diversity $C_s$ is given by the following:

$$C_s=-\sum _{i=1}^k\left({\displaystyle \frac{n_i}{N}}\right) log\left({\displaystyle \frac{n_i}{N}}\right)$$

(6)

For each color $C_i$, the connected regions within the image are calculated. $m_i$ denotes the number of connected regions of color $C_i$, and $r_{ij}$ denotes the number of pixels in the $j^{th}$ connected region. The spatial distribution complexity $C_{d,i}$ for the $i^{th}$ color is defined as follows:

$$C_{d,i}=-\sum _{j=1}^{m_i}\left({\displaystyle \frac{r_{ij}}{n_i}}\right) log\left({\displaystyle \frac{r_{ij}}{n_i}}\right)$$

(7)

To better calculate the differences between colors, the colors are converted from the HSV space to the perceptually uniform lab space. For each pair of colors $c_i$ and $c_j$, the Euclidean distance $D\left(c_i,c_j\right)$ in the lab space is computed and used as a weight for the complexity of the color distribution. Finally, the spatial distribution complexity is calculated for each color, and these values are weighted and aggregated to obtain the overall complexity of the color distribution $C_d$:

$$C_d=\sum _{i=1}^k\sum _{j=1}^k\left({\displaystyle \frac{n_i\cdot n_j}{N^2}}\right) D\left(c_i,c_j\right)\cdot C_{d,i}$$

(8)

where the Euclidean distance $D\left(c_i,c_j\right)$ is given by the following:

$$D\left(c_i,c_j\right)=\sqrt{{\left(a_i-a_j\right)}^2+{\left(b_i-b_j\right)}^2}$$

(9)

Here, $\left(a_i,b_i\right)$ and $\left(a_j,b_j\right)$ represent the coordinates of colors $c_i$ and $c_j$ in the lab color space.

This study utilizes only the a and b coordinates from the lab color space to calculate the distance between colors, aiming to minimize the interference of luminance variations and focus on the intrinsic differences among colors.

Compared to traditional color entropy methods, the spatial distribution complexity not only incorporates the distances between colors but also considers regional connectivity. Moreover, the HSV color model is more suitable for the adjustment or selection of color attributes, as it provides an intuitive means of manipulating hue, saturation, and brightness. In contrast, the lab color space, due to its perceptual uniformity, is better suited for quantifying visual differences between colors, especially when calculating color distances.

By employing Euclidean distance in the lab color space to quantify color differences and using these distances as weighting factors, the calculation of color complexity reflects not only the variety and frequency of colors but also their distribution patterns and adjacent relationships. Additionally, although converting colors between different color spaces may introduce minor discrepancies, these errors are generally negligible.

2.2.5. Grayscale Contrast

Grayscale contrast is a crucial metric for assessing local variations in grayscale values within an image. It reveals the visual impact of the image by evaluating the dynamic range of grayscale, the degree of polarization between black and white regions, edge sharpness, and the periodicity of repeating patterns [51]. Typically, images with high contrast reveal more intricate details and texture layers, offering richer visual information to the observer. Ciocca et al. used grayscale contrast as a metric for measuring image complexity, finding a notable correlation between the two [38]. In this study, we employed the skimage library in Python 3.10.0 to calculate grayscale contrast, primarily through the construction of a grayscale co-occurrence matrix (GLCM).

First, GLCM is generated by setting the distance between pixels and multiple directions (0°, 45°, 90°, and 135°) to represent the spatial relationships of pixel intensity values, capturing the variations in intensity at different angles. Next, contrast properties are extracted from these co-occurrence matrices, and the contrast values for each direction are averaged to obtain the overall grayscale contrast of the image.

The contrast is defined as the sum of the product of the squared intensity differences of each element in the co-occurrence matrix and their corresponding frequency of occurrence. The specific formula is as follows:

$$\mathrm{Contrast} =\sum _{i=0}^{N-1}\sum _{j=0}^{N-1}{(i-j)}^{2}P\left(i,j\right)$$

(10)

In this context, i represents the grayscale value of the first pixel, while j denotes the grayscale value of the adjacent pixel. $P\left(i,j\right)$ is an element in the co-occurrence matrix that indicates the probability of pixels with grayscale values of i and j co-occurring in the image at a specific direction and distance. The grayscale difference ${(i-j)}^2$ for each pair $\left(i-j\right)$ is multiplied by the co-occurrence probability $P\left(i,j\right)$, and the results are summed across all grayscale pairs. This process reflects the overall differences in grayscale values of adjacent pixels throughout the entire image.

2.2.6. Hierarchical Complexity

The aforementioned dimensions of complexity calculation focus on low-level image features, which, while significantly reflecting human visual complexity perception, still have limitations. Previous research has demonstrated that high-level semantic information is a critical factor influencing complexity [29,30]. However, existing studies rarely decode perceived complexity from high-level semantic information. This requires computers to mimic human understanding of images, identify various landscape elements, and quantify the diversity and richness of streetscape elements (such as buildings, street furniture, human activities, and signage).

In the research conducted by Jianping Fan et al., they proposed an innovative framework aimed at deeply exploring multi-level semantic information in images through a hierarchical classification approach. This method effectively narrows the gap between low-level visual features and human high-level semantic understanding of images. In this framework, an image is structured as a tree, with each node representing a specific image concept or prominent object category, forming a hierarchical semantic representation [31].

In image compression, tree-structured vector quantization techniques are used to calculate node depth and quantity based on feature data within the image, which can partially reflect the image’s visual complexity [52]. Inspired by this theory, our study combines these two approaches to propose a novel method—hierarchical complexity—to measure the richness and diversity of landscape objects in streetscape images.

The core of hierarchical complexity lies in the precise semantic segmentation achieved using the Deeplabv3+ model. To ensure that the model can adapt to the complexities of real-world street scenes, we selected the Cityscapes dataset for model training and testing. Compared to other datasets, the Cityscapes dataset is a high-quality resource specifically designed for urban streetscape analysis, encompassing 19 classes of visual element labels that include the most significant visual components found in street environments. Additionally, the moderate number of classes in Cityscapes helps to reduce model complexity, facilitating easier training and optimization of the semantic segmentation model.

Furthermore, Deeplabv3+ has demonstrated outstanding performance when processing urban streetscape datasets (Cityscapes) due to its ability to capture multi-scale information within images. In this study, we opted for the Deeplabv3+ model with ResNet-18 as the backbone network, striking a balance between computational efficiency and model performance.

During the semantic segmentation process, 19 different elements will be comprehensively incorporated into the semantic hierarchy. This includes movable elements such as pedestrians and vehicles, which are not only critical components of street scenes but also influence individuals’ subjective perceptions and psychological experiences on multiple levels [53,54]. These elements can objectively reflect the actual usage conditions of the streets, thereby impacting the perception of street scene complexity.

After segmentation, as in Figure 5, each minimal street element is labeled, and corresponding nodes are created in the image’s semantic tree structure. By analyzing the inclusion relationships among these nodes, a semantic tree is constructed, reflecting the hierarchy and interrelationships of the objects. The segmented image elements are first categorized into 19 subclasses, which may represent specific segmentation labels such as road, building, vehicle, and so on. These subclasses are further aggregated into six major categories: roads, pedestrians and vehicles, vegetation, traffic facilities, buildings, and others. Each major category is assigned a different weight based on its proportional presence in the image, with the weight distribution determined by its influence on the complexity of the image’s hierarchical structure. We quantify the number of nodes in this tree structure and incorporate these varying weights to reflect the hierarchical complexity of the image.

Specifically, the hierarchical complexity is calculated using three weighted factors: the number of levels $H$, the number of weighted elements $W_E$, and the number of objects $O$. The formula is as follows:

$$Hierarchical Complexity=H_w\cdot H+E_w\cdot W_E+O_w\cdot O$$

(11)

where $H_w$, $E_w$, and $O_w$ are the respective weight values for the levels, weighted elements, and objects. The number of objects $O$ at the first level of the tree structure is obtained by counting the contours corresponding to each segmentation label, as represented by the following formula:

$$O=\sum _{j=0}^M{L}_j$$

(12)

where $L_j$ is the number of contours corresponding to each element label, and $M$ is the total number of element labels present in the image. This formula reflects the cumulative count of objects for each label in the image.

At the second level of the tree structure, the element counts of different major categories are calculated using a weighted approach, as shown in the following formula:

$$W_E=\sum _{k=1}^P{W}_k\cdot C_k$$

(13)

where $W_k$ represents the weight value for different categories, and $C_k$ is the number of objects for each label. The weight for each category is allocated based on its relative importance within the image’s hierarchical structure. The effect is illustrated in the accompanying Figure 6.

2.3. SVM-Trained Model

2.3.1. Model Training Label

In this study, expert evaluation was employed to score the training sample images used as training labels for the SVM model. A total of 15 participants, including students and professors specializing in landscape architecture, were invited to assess the complexity of 560 randomly selected streetscape images from the study area. Visual complexity was categorized into three levels: low, medium, and high, corresponding to scores of one, two, and three, respectively. To ensure the reliability and consistency of the scores, a subset of images was randomly selected from the 560 samples for a preliminary joint evaluation by all experts to reach a consensus on the scoring criteria. Following this consensus, each expert independently evaluated the complexity of the 560 images. After completing the evaluation, the average complexity score for each image was calculated, rounded to the nearest whole number, and used as the training label for the SVM model.

2.3.2. Building the SVM-Trained Model

The SVM is a supervised learning algorithm widely used for classification and regression tasks. It identifies the optimal separating hyperplane by maximizing the width of the decision boundary, demonstrating robustness to noise and outliers, and exhibiting unique advantages in high-dimensional spaces [55,56,57]. SVM has been repeatedly applied in studies quantifying visual complexity [58,59]. Before inputting data into the SVM classification model, six metrics—compression ratio, symmetry, fractal complexity, color complexity, grayscale contrast, and hierarchical complexity—were calculated for each of the 560 training sample images. The distribution of these sample data is visualized through box plots and kernel density estimates, as shown in Figure 7a. After normalizing these six metrics, they were used as input features for model training, with the expert-assessed visual complexity serving as the training labels. During the model training phase, we meticulously tuned the parameters of the SVM model to optimize its performance. Following a series of experiments and validations, we decided to utilize a linear kernel function as the kernel for the SVM, as it demonstrated superior classification effectiveness.

2.3.3. Model Validation

A correlation analysis of the input features was conducted, and a heatmap of the feature matrix was created, as shown in Figure 7b. The results indicate that there is a moderate to low correlation among the features, with no significant linear dependencies observed. This suggests that the six metrics—compression ratio, symmetry, fractal complexity, color complexity, gray level contrast, and hierarchical complexity—each represent distinct dimensions of complexity. Each metric provides unique information, ensuring diversity and complementarity among the model’s input features and offering a comprehensive informational basis for the model.

To validate and assess the accuracy of the trained SVM model, a fit analysis was conducted using a five-fold cross-validation method. This approach involved dividing the dataset into five subsets, with each subset serving as the test set once while the remaining four subsets were used for training. This methodology ensured a comprehensive and reliable evaluation of the model. The results indicated that the SVM classification model demonstrated a high level of consistency in capturing complexity perceptions, achieving an average accuracy of 84.05%. This performance highlights the model’s potential for practical applications, reflecting its ability to accurately classify data and effectively understand and represent various dimensions of complexity.

3. Results

3.1. Contribution of Features in the SVM Classification Model

The contribution of each feature in the SVM classification model is typically determined by the model parameters obtained during training, which reflect the extent of each feature’s contribution to the model. The features are ranked by their contribution in the following order: compression ratio, grayscale contrast, hierarchical complexity, fractal dimension, color complexity, and symmetry (

Table 1. Percentage contribution of each feature.

Compression Ratio	Symmetry	Color Complexity	Grayscale Contrast	Fractal Dimension	Hierarchy Complexity
29.02%	2.53%	6.65%	26.69%	14.27%	20.84%

).

Despite its proven close relationship with visual complexity, the compression ratio has rarely been used as a metric to quantify image complexity in previous studies. In this paper, the compression ratio is employed to measure visual complexity and is identified as the feature with the highest contribution. This validates the strong connection between the compression ratio and complexity perception. Experimental results show that images with a higher compression ratio often contain extensive details and complex structures, such as intricate building textures and plant details in street scenes. These images are difficult to compress without significant loss of information, resulting in minimal differences in file size before and after compression.

High grayscale contrast areas, characterized by their distinct light-dark distinctions, tend to capture greater attention, thereby enhancing the resolution of visual details and improving the perception of three-dimensionality in the image [60]. As a result, grayscale contrast is particularly crucial in evaluating the complexity of visual scenes. In street scenes, certain elements exhibit high grayscale contrast due to significant brightness differences from their surroundings, while most buildings or ground structures show more gradual brightness changes, resulting in lower grayscale contrast.

Hierarchical complexity demonstrates a strong performance in the SVM model’s contribution analysis. It reflects the diversity of the street and offers insights into its vibrancy, pedestrian activity, and traffic volume. This complexity assessment, which utilizes advanced image semantic segmentation techniques, enhances traditional methods by capturing often-overlooked details. By leveraging deep learning on extensive streetscape datasets, semantic segmentation models can accurately extract high-level semantic information from images, thereby providing a more nuanced simulation of human perception of complexity.

A higher fractal dimension indicates greater self-similarity and more complex details across different scales [46]. Streetscape images often contain numerous self-similar structures, such as the branching of trees, the arrangement of building windows, and the patterns in brick walls. They reflect the visual complexity and rich detail layers of street scene images and align well with subjective perceptions of complexity.

Previous research has suggested that visual features like color are key factors in increasing landscape diversity and significantly impact complexity perception in urban environments [61,62]. However, Ciocca et al. (2015) presented a contrasting view, finding that metrics such as color richness do not correlate as strongly with complexity perception as expected [9]. This paper also explores the impact of color complexity on complexity perception and finds that while color complexity plays a role, its influence is relatively limited. In street scene image analysis, despite the rich visual information provided by color variation, grayscale information may more accurately reflect image complexity in certain contexts [38]. For instance, under low light or adverse weather conditions, color information in street scene images may be diminished, whereas grayscale contrast and structural information maintain higher stability.

Symmetry contributes the least to the model, providing minimal information on complexity and not being a primary evaluative feature. Common symmetrical structures in street scenes include the symmetrical design of buildings and the mirrored layout of roads. However, these symmetrical elements often do not dominate the overall image, especially in scenes featuring complex natural landscapes. Street scene images frequently contain dynamic elements, such as pedestrians and vehicles, whose positions and shapes vary randomly, increasing the image’s asymmetry and complexity. Features like compression ratio, grayscale contrast, fractal dimension, and hierarchical complexity better capture the details and multi-scale information of images, thus more effectively reflecting the complexity of street scene images.

Although color complexity and symmetry contribute less to the model’s performance, they still marginally improve its accuracy. Therefore, these two features were retained in the current model. Additionally, the model developed in this study exhibits strong generalization capabilities. In future applications, when computational resources are limited or model simplification is necessary, omitting the color complexity and symmetry features could optimize computational efficiency without significantly impacting the model’s overall performance.

3.2. Practical Application of the SVM Classification Model

Through an in-depth analysis of 1800 coordinate points in the Xiaobailou and Wudadao Districts of Tianjin, we extracted six key features from 6844 street scene images: compression ratio, symmetry, color complexity, grayscale contrast, fractal dimension, and hierarchical complexity. These six features were input into a trained SVM classification model to predict image complexity perception.

The results indicated that images rated one point typically exhibited a lack of street features and homogeneity, with monotonous color composition and often large areas of blank space. In contrast, images rated three points demonstrated rich landscape features and diverse elements, densely distributed throughout the image.

Images rated with a score of two exhibit moderate complexity, typically containing a balanced amount of landscape features, falling between simplicity and high complexity. Buildings in these images often display some degree of detail but are neither overly dense nor highly intricate, usually showing a uniform and orderly distribution. The streetscape includes basic greenery elements, such as trees along the roadside, small greenbelts, or flowerbeds. Additionally, pedestrians and vehicles (e.g., cars, bicycles) appear in moderate numbers without causing a sense of overcrowding or chaos. Overall, these images maintain visual balance, offering enough detail to capture interest without overwhelming the viewer, resulting in a harmonious and moderately complex visual experience.

The visual complexity predictions for each streetscape image were mapped to their respective geographic coordinates. Each coordinate point had four different streetscape images, so the average complexity for each coordinate point was calculated. This approach effectively mitigates potential biases that could arise from single-direction images, ensuring a more comprehensive and objective assessment of the complexity of the street environment. As Figure 8 shows, the visual complexity scores for the Xiaobailou and Wudadao Districts in Tianjin ranged from one to three, with colors varying from dark to light. Areas in the central and southwestern parts of the map showed lighter colors of complexity, corresponding to higher scores, indicating greater complexity. These regions exhibit higher land development intensity, predominantly comprising residential and commercial areas. The variety of buildings—including low-rise and high-rise residential structures as well as office towers—combined with well-maintained greenery contributes to a richer and more diverse landscape. Additionally, the high pedestrian and vehicular traffic in these areas likely contributes to the increased complexity of the street scenes.

Conversely, areas with darker colors located along the edges of the map showed lower visual complexity. These areas generally have moderate land development intensity, shorter buildings, fewer detailed structures, and lower building density. For instance, the Wudadao Cultural Tourism District, located in the southwestern part of the map, primarily features low-rise buildings, including detached and row houses, with a floor area ratio mostly less than 1. These buildings, inspired by classical revival, romanticism, and eclectic styles popular in Europe and the United States, are also infused with traditional Chinese architectural elements, resulting in a more minimalist overall style [63]. Additionally, the lower skyline of these low-rise buildings provides open vistas and comfortable living spaces, which may contribute to the lower complexity scores in image feature analysis.

3.3. Distribution of Image Features in the Study Area

To better visualize the geographic distribution of different image feature data and to gain a detailed understanding of various dimensions of visual complexity across certain streets in Tianjin, we conducted a geographic mapping of these six features. The value at each latitude and longitude point represents the average score of street views in different directions at that location.

Among the features, the visual distribution of compression ratio (Figure 9a), hierarchical complexity (Figure 9b), and fractal dimension (Figure 10a) appear quite similar. The darker-colored regions on the maps are primarily concentrated in the northern and central parts. These areas mainly cover the Xiaobailou District, which is home to numerous historical buildings such as the former French Consulate, the former Chosun Bank, and the former Beiyang Commercial Bank. These buildings, mostly constructed in the 1930s, showcase the essence of modern Western architecture, including Baroque and Rococo styles [64]. The Xiaobailou District is a vibrant commercial hub characterized by high foot traffic. The district’s elaborate storefronts and distinctive signage contribute to its unique visual and cultural landscape, resulting in high complexity scores for hierarchical complexity, compression ratio, and fractal dimension.

In contrast, the lighter-colored regions are mainly located in the lower and peripheral areas of the map, encompassing the Wudadao Cultural Tourism Area. This area features low-rise buildings, predominantly standalone and row villas, offering relatively open views. The lack of detailed landscape elements and the more uniform vegetation in these regions likely contribute to the lower complexity scores in compression ratio, hierarchical complexity, and fractal dimension.

In the analysis of grayscale contrast (Figure 10b), we observed a broad range of score fluctuations, ranging from 500 to 2500. Significant differences in grayscale contrast were found across various areas. High-contrast points are scattered throughout the map but are particularly concentrated in certain areas of Liuzhou Road, Heilongjiang Road, and Binjiang Road. In contrast, areas with lower grayscale contrast are more evenly distributed across the map. This uniformity likely reflects consistent building materials and design or stable lighting conditions in these areas, leading to overall lower grayscale contrast.

The color complexity (Figure 11a) mapping indicates that the central and southern areas of the Wudadao District display greater color complexity, which accurately reflects the current conditions. The architectural styles in these areas are highly diverse, with unique variations in the color combinations of buildings and other visual elements such as vegetation. The Wudadao District is a prominent tourist area in Tianjin, rich in cultural elements, with extensive use of color in architectural carvings, signage, decorations, and advertisements. These factors collectively contribute to the high scores in color complexity in this region.

In terms of symmetry (Figure 11b) scores, locations with higher scores are predominantly concentrated in the Wudadao District. This may be attributed to the architectural style of the Wudadao District, which tends to feature lower-rise buildings and more uniform architecture on both sides of the streets. Additionally, Dagu North Road also exhibits notable symmetry. However, Dagu North Road scores poorly in color complexity and grayscale contrast. This may be related to the modern architectural style of high-rise office buildings on either side of the road, which tends toward uniformity in design, resulting in a highly consistent visual experience on both sides of the street. Since the evaluation of symmetry primarily relies on the analysis of grayscale histograms, symmetry scores are naturally higher when the color distribution on both sides of the street is similar. Therefore, Dagu North Road scores high in symmetry, while its scores for grayscale contrast and color complexity are relatively low.

Visual complexity arises from the interplay of multiple visual features, with each feature contributing differently to the overall complexity. Taking the Wudadao District as an example, it performs well in color complexity and symmetry but is relatively average in compression ratio, grayscale contrast, hierarchical complexity, and fractal dimension. Nonetheless, these four image features contribute to over 90% of the overall visual complexity, leading to a lower overall complexity score for the Wudadao District.

3.4. Complexity Scores and Typical Streets in the Study Area

3.4.1. Complexity Scores of Streetscapes

The streetscape of Tianjin is renowned for its diverse characteristics and the fusion of historical and modern elements. This study focused on the Xiaobailou District and the Wudadao District, encompassing 76 streets of various classifications, including major roads, secondary roads, and alleys. To ensure the reliability and validity of the research results, a rigorous selection process was employed. Streets with fewer than five sampling points or fewer than 20 street view images were excluded, as their shorter lengths and insufficient sample sizes would not provide adequate statistical representativeness and analytical value. After this selection process, a total of 57 streets were included in the study. The average complexity scores of these streets are shown in Figure 12, while the distribution of complexity levels across different streets is depicted in Figure 13.

Through a detailed analysis of Figure 12 and Figure 13, we identified significant variations in visual complexity levels among the streets within the study area. As depicted in Figure 12, the average complexity scores for most streets range between 1.6 and 2.4, suggesting that the streetscapes generally display a moderate level of visual complexity. Additionally, Figure 13 shows that images with a complexity score of two are the most prevalent across the majority of streets. This observation corroborates the findings from Figure 12, reinforcing the conclusion that the overall visual complexity of the streetscapes in the study area is moderate.

3.4.2. Moderate Complexity Streets

As illustrated in Figure 12 and Figure 13, the visual complexity of the streetscape in the Xiaobailou and Wudadao Districts predominantly remains at a moderate level. Shanxi Road, Qufu Road, and Hebei Road are typical examples of streets with moderate complexity within the study area. These streets not only represent the historic districts of Tianjin, reflecting the city’s rich historical and cultural characteristics but also hold significant commercial value. Figure 14 shows the basic distribution of street views, revealing distinct visual styles among the three roads. Shanxi Road and Hebei Road, being one-way streets, are relatively narrow, with predominantly residential buildings that are low-rise and inclined toward Western classical architectural styles. In contrast, Qufu Road, as a secondary arterial road, is wider and more accessible, with modern high-rise buildings. The color-coded annotations in the figure represent different complexity levels, with a central concentration of moderate complexity and fewer at the extremes. Most street views exhibit moderate complexity, contributing to a balanced, appealing urban environment.

3.4.3. High-Complexity Streets

Although most streets in the study area exhibit moderate visual complexity, certain streets significantly exceed this level. As shown in Figure 13, Yueyang Ave, Yantai Road, Xi’an Ave, Shashi Ave, Liuzhou Road, and Dalian Road have streetscape images where the complexity score of three exceeds 40%. Figure 12 and Figure 13 indicate that Liuzhou Road has the highest average visual complexity, with over 70% of the images scoring three. As in Figure 15, there are only three images with a score of one and six images with a score of two, while the majority received the highest rating of three. Although some studies suggest that excessively complex built environments are less favored in building environments [65], Many scholars have noted that the relationship between complexity and preference is not straightforward. Seemingly chaotic environments might lack consistency, but highly consistent landscapes can still positively influence environmental preferences even if they are complex [23,66].

Liuzhou Road, as shown, features a rich array of visual elements and diverse landscape characteristics. The area is densely populated with commercial elements, resulting in a rich green layer and a well-defined skyline. The streets are lined with commercial advertisements and signs, creating visual contrasts and rhythms between distant high-rise buildings and nearby low-rise structures, which adds depth and complexity to the landscape. High pedestrian and vehicle traffic contributes to a rich visual layer with compact spatial usage within the district. Despite its high complexity, the street maintains a degree of consistency, helping to avoid visual confusion.

3.4.4. Low-Complexity Streets

Conversely, the streets Taierzhuang Road, Nanning Road, Jiefang North Road, Hejiang Road, and Datong Road, including Baoding Bridge, have a proportion of street view images with a complexity score of one exceeding 40%. These streets show relatively weak performance in terms of visual diversity and richness. For instance, Nanning Road (Figure 16), with the lowest average visual complexity in the study area, has most of its street view images scoring between one and two. The overall streetscape of Nanning Road is characterized by uniformity and lack of appeal. The low-rise buildings, primarily mid-to-low-rise, feature simple facades and minimal greenery, failing to meet the visual preferences of nearby residents, which may affect overall environmental satisfaction. Additionally, low-complexity streets do not provide sufficient sensory stimulation, potentially leading to a lack of psychological relaxation and enjoyment [45]. Future planning and optimization are necessary to enhance the overall landscape effect and improve residents’ experience.

4. Discussion

4.1. SVM Quantification Model

Visual complexity is a crucial dimension in theories such as urban quality assessment [26], arousal theory [21], and environmental preference theory [67]. Traditional assessment methods rely heavily on subjective evaluations, which present several challenges. First, they involve high time costs, as evaluators must invest considerable effort in researching and analyzing subjects to establish standardized criteria. Second, the scope of evaluation is limited, with manual assessments typically confined to small-scale images and exhibiting poor generalizability. Lastly, there is a feedback delay issue, where rapidly changing environments can quickly render evaluations obsolete.

To address these problems, this study developed an objective and reliable visual complexity quantification model by training the mapping relationship between image features and human subjective perception. We employed the SVM as the core model and selected six key image features as inputs: compression ratio, symmetry, fractal dimension, color complexity, grayscale contrast, and hierarchical complexity. These features encompass both high-level semantic understanding of images and low-level feature computations, offering a comprehensive perspective on quantifying visual complexity in landscapes.

The SVM visual complexity quantification model demonstrates significant advantages, achieving an accuracy of 84.05%. It efficiently and accurately predicts visual complexity in large-scale streetscape data using a small sample for training. Compared to existing complexity quantification methods, this study innovatively introduces the concept of hierarchical complexity, specifically designed to quantify visual richness in landscapes. This approach effectively simulates human perceptual processes, bridging the gap between high-level semantic understanding and low-level image features, thus covering a broader range of dimensions in visual complexity.

The hierarchical complexity is based on the Cityscapes dataset, which segments streetscape images into 19 labeled categories. Although real-world streetscapes feature various types of buildings and vegetation, using hierarchical complexity alone to directly reflect the diversity of these elements has certain limitations. This is primarily due to the reliance of hierarchical complexity on semantic segmentation techniques and the limited availability of mature computer vision models suitable for this task. Additionally, constraints in computational resources further increase the difficulty. However, other complexity features introduced in this study effectively compensate for these limitations. For instance, fractal dimension measures the complexity of shapes, color complexity captures the richness of hues, compression ratio indirectly reveals the density of information, and grayscale contrast captures detailed variations within the image. By integrating these features for analysis, we can achieve a more comprehensive assessment of the overall visual complexity of street scene images despite the inherent limitations of hierarchical complexity.

Furthermore, this study focused on the Xiaobailou and Wudadao Districts in Tianjin to predict and analyze the visual complexity of streets in these regions while validating the effectiveness of the SVM quantification model. The results indicate that most streets in the study area exhibit moderate visual complexity. High- and low-complexity streetscapes show distinct visual differences, which can be broadly categorized into three aspects: pedestrian and vehicular traffic volume, the richness of architectural façade decorations, and the layering and arrangement of vegetation. Streets with high visual complexity are typically associated with higher development intensity, featuring abundant commercial facilities and diverse landscape structures, providing a richer and more dynamic visual experience. This finding aligns with existing knowledge. Conversely, streets with low visual complexity exhibit lower development levels, with relatively uniform visual effects and lack of appeal.

To further validate the effectiveness of the evaluation results in this study, we compared them with the street conditions reported in the “Tianjin Heping District 2023 Government Work Report” [68]. We found that the visual complexity scores from our study are consistent with the actual street conditions. Streets highlighted in the report as needing improvement, such as Binjiang Road and Hebei Road, have an average visual complexity score slightly above two. This reflects issues with the streetscape, such as inconsistent styles of street-facing shops and outdated residential buildings, leading to high visual complexity but also visual disorder. Future efforts should focus on enhancing the development and updating of these streets, continually optimizing spatial layouts to further enhance their visual appeal.

4.2. Analysis of Streetscape Complexity Features

Visual complexity is a crucial explanatory factor influencing environmental preferences [7]. It directly affects the visual quality assessment of streetscapes and, in turn, impacts residents’ satisfaction with their environment. This study identifies several factors that shape streetscapes’ visual complexity: compression ratio, grayscale contrast, hierarchical complexity, fractal dimension, color complexity, and symmetry. Notably, compression ratio, grayscale contrast, hierarchical complexity, and fractal dimension have particularly significant effects on street visual complexity. When designing street views, evaluating performance across these dimensions and adjusting image features accordingly can help strategically enhance or reduce visual complexity.

Specifically, compression ratio and grayscale contrast primarily affect image detail, which can be adjusted by modifying texture or fine structural elements. Hierarchical complexity is closely related to the layering and diversity of the environment, which can be altered by increasing or decreasing the layout of streetscape elements and plants. The fractal dimension is linked to the richness of natural elements, which can be adjusted by varying the distribution of trees and grass to enhance the natural aesthetics of the street view. Detailed analysis and adjustments based on these findings can ensure that street designs are visually attractive while harmoniously integrating into their environment.

4.3. Strategies for Optimizing Street Visuals

In the study area, most streetscapes exhibit a medium level of complexity, with some streetscapes showing high or low complexity. It is widely believed that environments with moderate complexity are aesthetically pleasing [69,70], as they provide a balanced amount of information and avoid monotony or information overload. The study reveals that there is an optimal cognitive load level when processing visual information. Overly simplified landscapes may lack appeal, often leading to feelings of boredom and depression [19], while excessively complex landscapes can cause visual fatigue and lead to information overload [21]. Medium-complexity environments achieve visual balance and harmony, stimulating brain activity without causing stress, which is more likely to induce pleasure. Alpak’s research further confirms that people exhibit a marked preference for urban built environments with moderate complexity [67]. These environments not only resonate with residents’ visual aesthetics and environmental preferences but also have a positive impact on their psychological well-being. To further enhance the visual appeal and cultural atmosphere of streets with moderate complexity, future efforts could focus on regular maintenance of plantings and building facades, the introduction of vertical greenery, and the incorporation of public art installations, thereby elevating their overall aesthetic value.

High-complexity streetscapes are closely related to the volume of information they convey, presenting greater challenges for design. During the design process, it is essential to meticulously coordinate various elements of the environment and employ a layered design approach, which includes the organized arrangement of the foreground, middle ground, and background. By skillfully integrating architectural features and selecting appropriate paving materials, color schemes, and plants of varying heights, designers can maintain overall visual richness while avoiding visual clutter.

Low-complexity streetscapes often suffer from insufficient development, leading to minimal greenery and beautification, monotonous architectural styles, a lack of commercial facilities, and inadequate infrastructure, which can also affect foot traffic. Design strategies should focus on adding detail to increase visual complexity. For example, incorporating diverse commercial signage systems and integrated service facilities, as well as creating visual depth with varied plant configurations and decorative patterns on building facades, can attract visual interest and enhance the overall appeal of the environment.

In urban environment design, the visual complexity of streetscapes is a key factor affecting their attractiveness and harmony. Adjusting street designs to balance visual complexity can optimize the visual effects of streetscape. This approach helps maintain street attractiveness while avoiding excessive complexity or monotony, thereby creating aesthetically pleasing and harmonious street environments.

4.4. Limitations and Future Directions

This study has some limitations. Although the proposed streetscape complexity quantification model predicts visual complexity with an accuracy of 84.05%, there remains some discrepancy compared to human subjective perception. This discrepancy may arise from the following reasons. (1) Visual complexity dimensions: Human subjective perception of complexity encompasses a more diverse range of dimensions than the indicators proposed in this study, which may not fully cover all aspects. (2) Technological limitations: current computer vision technologies are not yet fully mature, and the accuracy of algorithms for quantifying complexity indicators needs improvement. For instance, image semantic segmentation, required for calculating hierarchical complexity, is inconsistent across different types and scenes, affecting the precise calculation of complexity indicators. (3) Image conditions: despite preprocessing, variations in image capture conditions—such as lighting, exposure, and focal length—can affect calculation accuracy. For example, lighting changes can influence edge detection, and inadequate or excessive exposure can lead to increased noise or loss of detail.

Despite numerous studies confirming a close relationship between perceived complexity and theories of environmental quality assessment, environmental preferences, arousal theory, and restorative benefits, a unified definition of these relationships has not yet been established due to the challenges of quantifying complexity and technological limitations. This research aims to provide a more scientific quantitative approach for future studies, facilitating a deeper exploration of these theoretical connections. Additionally, differences in image features and human environmental preferences have been observed, suggesting that future research could further investigate this direction.

Moreover, the model training samples in this study were primarily sourced from the Xiaobailou and Wudadao Districts, which may introduce some regional heterogeneity in capturing area-specific features. There may be significant differences in the numerical distribution of streetscape image features across different regions. Therefore, for future studies using this model to predict streetscape complexity in other areas, it is recommended to use local street view images as training samples to achieve more accurate predictions of perceived complexity.

5. Conclusions

This research aims to apply artificial intelligence technology to develop an objective quantification model for streetscape complexity perception, addressing limitations in traditional visual complexity assessments such as scope restrictions, data acquisition challenges, and measurement accuracy. The Xiaobailou and Wudadao Districts in Tianjin were selected as the study area. Streetscape images were collected using the Baidu Maps API, with a small subset of images used for model training. The trained model was then applied to evaluate the complexity perception of the remaining images. The main findings of the study are the following:

• This study introduces an innovative dimension for measuring visual complexity—hierarchical complexity. This dimension serves as a tool designed specifically for assessing landscape visual complexity and bridges the gap between low-level semantic features in streetscape images and high-level human semantic cognition. Hierarchical complexity is based on high-level semantic information from images, allowing for a more comprehensive capture of visual complexity in streetscapes and reflecting the richness of streetscape elements. The DeepLabv3+ model was utilized for precise semantic segmentation of streetscape images, marking the smallest environment elements and creating corresponding nodes in the image’s semantic tree structure. By analyzing the inclusion relationships between these nodes, a hierarchical semantic tree was constructed, with the number of nodes in the hierarchy serving as an indicator of hierarchical complexity.
• This study developed a SVM-based model for quantifying streetscape visual complexity, achieving an average accuracy rate of 84.05%. This model can automate the processing of large volumes of image data, significantly improving efficiency and reducing the influence of subjective factors, thereby ensuring consistency and objectivity in measurements. The SVM model incorporates six input features across multiple dimensions, encompassing both low-level image features and high-level semantic information. The contribution of these features, ranked from highest to lowest, is as follows: compression ratio, grayscale contrast, hierarchical structural complexity, fractal dimension, color complexity, and symmetry. Among these, compression ratio, grayscale contrast, hierarchical structural complexity, and fractal dimension are the key factors influencing visual complexity. This ranking facilitates a more precise identification and understanding of each feature’s role in streetscape visual complexity. This, in turn, allows urban planners and environmental designers to tailor street designs and optimize visual complexity, thereby improving both the aesthetic appeal and functional quality of urban spaces.
• The quantification model for streetscape visual complexity was applied to the Xiaobailou and Wudadao Districts in Tianjin. Detailed statistical and analytical evaluations were conducted for each street within the study area. Most streets exhibited moderate levels of visual complexity, with fewer streets falling into low- or high-complexity categories. High-complexity streetscapes typically featured diverse building types, rich vegetation, and high foot and vehicle traffic, such as in Yueyang Ave, Yantai Road, Xi’an Ave, Shashi Road, Liuzhou Road, and Dalian Road. Conversely, low-complexity street views often had uniform landscape elements and large blank areas, as seen in Taierzhuang Road, Nanning Road, Jiefang North Road, Hejiang Road, Datong Road, and Baoding Bridge. Based on the typical streets of different complexity levels, corresponding improvement recommendations have been proposed.

This study successfully developed a quantification model for urban streetscape visual complexity, offering a more scientific and precise tool for future research. It also emphasizes the importance of visual complexity in urban space design, offering a new perspective on balancing urban aesthetics and functionality. We hope that the development of this model will advance in-depth discussions in the fields of built environment and visual perception, providing urban planners and environmental designers with a robust data foundation. This will enable more precise assessment and adjustment of visual complexity in urban spaces, ultimately optimizing urban environments, enhancing residents’ quality of life, and elevating the aesthetic value of cities.