Enhancing Dongba Pictograph Recognition Using Convolutional Neural Networks and Data Augmentation Techniques

Abstract

The recognition of Dongba pictographs presents significant challenges due to the pitfalls in traditional feature extraction methods, classification algorithms’ high complexity, and generalization ability. This study proposes a convolutional neural network (CNN)-based image classification method to enhance the accuracy and efficiency of Dongba pictograph recognition. The research begins with collecting and manually categorizing Dongba pictograph images, followed by these preprocessing steps to improve image quality: normalization, grayscale conversion, filtering, denoising, and binarization. The dataset, comprising 70,000 image samples, is categorized into 18 classes based on shape characteristics and manual annotations. A CNN model is then trained using a dataset that is split into training (with 70% of all the samples), validation (20%), and test (10%) sets. In particular, data augmentation techniques, including rotation, affine transformation, scaling, and translation, are applied to enhance classification accuracy. Experimental results demonstrate that the proposed model achieves a classification accuracy of 99.43% and consistently outperforms other conventional methods, with its performance peaking at 99.84% under optimized training conditions—specifically, with 75 training epochs and a batch size of 512. This study provides a robust and efficient solution for automatically classifying Dongba pictographs, contributing to their digital preservation and scholarly research. By leveraging deep learning techniques, the proposed approach facilitates the rapid and precise identification of Dongba hieroglyphs, supporting the ongoing efforts in cultural heritage preservation and the broader application of artificial intelligence in linguistic studies.

1. Introduction

In recent years, artificial intelligence (AI) has transformed numerous domains, yet the integration of its advanced techniques into heritage preservation remains uneven. Deep learning, particularly Convolutional Neural Networks (CNNs), has demonstrated remarkable success in visual recognition tasks, but its adoption in traditional village landscape analysis and architectural heritage documentation is still emerging. The increasing need for efficient, scalable, and accurate tools to safeguard intangible cultural assets raises a critical question: how can AI-driven models be leveraged to enhance the classification and retrieval of heritage imagery? Prior research has shown the potential of CNNs in heritage contexts. For example, Wang et al. [1] proposed a pixel-level semantic segmentation approach to identify spatial characteristics in traditional village landscapes. While effective in simulating expert perception, this model lacked adaptability across diverse datasets. Similarly, Gao et al. [2] introduced a two-stage CNN Attention Retrieval Framework (CNNAR), combining transfer learning with attention mechanisms to classify Chinese diaspora architectural images. Although their model achieved high retrieval accuracy, it primarily focused on narrowing the search scope rather than enhancing interpretability across various image types.

Other studies have further highlighted the trade-offs between performance and flexibility in heritage-related AI applications. Janković [3] compared neural and non-neural models using handcrafted features, revealing that while multilayer perceptron models excelled in efficiency, they were limited in scalability and adaptability. Ju [4], through a macro-level bibliometric review, revealed an evolving trend in AI-for-heritage research—one that increasingly intersects with sustainability, material degradation, and ecological monitoring. Parallel developments in dataset design have supported this shift: the ImageOP dataset Version 1 [5] enabled low-resource image classification on mobile devices, while Go et al. [6] compared CNNs and Vision Transformers (ViTs) in pigment classification, finding ViTs highly accurate but less interpretable. Despite these advancements, a critical gap persists: few models integrate high classification accuracy, real-time retrieval, and visual interpretability across culturally diverse datasets.

The Dongba script, a unique pictographic writing system used by the Naxi people in Yunnan Province, China, represents one of the few surviving pictographic scripts in the world. Despite its historical significance and cultural richness, the script has faced challenges with modernization, preservation, and digital integration. Over the years, scholars have conducted extensive research to analyze its grammatological features, structural evolution, and practical applications in various domains, ranging from linguistic studies to digital humanities and virtual reality applications.

The systematic study of the Dongba pictographic script has evolved through the application of digital tools and network analysis approaches. Xu [7] describes implementing several digital tools to facilitate grammatological studies, emphasizing the importance of a structured dictionary and an online database of Dongba characters. By organizing entries based on major semantic components, the database allows for statistical analysis, visualization of glyph clusters, and a closer examination of individual glyph elements. Furthermore, network analysis techniques provide insights into the developmental stages of the writing system, highlighting the structural complexity and interrelationships among its components.

The visual and symbolic characteristics of Dongba pictographs have also influenced modern pictogram design. Zhen [8] explores new methods for pictogram development inspired by Dongba formative principles. These methods were tested through Traditional Chinese Medicine pictograms and assessed for readability via questionnaire-based evaluations. Similarly, the “Dongba Script Character Construction Space” utilizes virtual reality technology to create an immersive experience that enables interaction with pictographic, ideographic, and compound ideographic forms [9]. Such efforts demonstrate the potential for preserving and revitalizing the Dongba script through innovative digital applications.

Historical documentation further enriches the understanding of the script’s evolution. The woodblock prints of the “Naxi Pictographs and Geba Script Comparative Lexicon”, housed at the Lijiang Museum, are crucial artifacts reflecting the Naxi script unification movement of the early 20th century [10]. These prints provide evidence of script standardization efforts by Dongba ritualists, shedding light on cultural and historical factors that influenced the development and decline of the unified script.

Phonological research has also contributed to the study of the Dongba script, particularly in its relationship with the Naxi language. Xu [11] examines the use of Dongba glyphs in representing phonetic structures, noting their efficiency in identifying initials but their limitations in distinguishing rhymes and tones. A phonemic chart written in Dongba glyphs serves as a tool for analyzing the phonological system and tracking sound changes in Ruke Naxi. This study underscores the adaptability of the Dongba script in linguistic research and its potential for phonetic representation.

In the field of computational linguistics, automated analysis of the Dongba script has gained attention. Text line segmentation is a fundamental preprocessing step for Dongba document analysis, facilitating tasks such as character extraction, translation, and annotation. Yang and Kang [12] propose a segmentation algorithm combining K-means clustering with projection-based methods to improve text block recognition and extraction. Similar computational approaches are applied to the structural analysis of Dongba hieroglyphs, using the Hough transform algorithm for table recognition and database classification [13]. Such advancements provide a technical foundation for further digital research and preservation efforts.

Ritualistic and cultural aspects of the Dongba script are also significant. He and Zhao [14] analyze the “eye disease incantation”, a unique category of Dongba manuscripts that combines textual, pictographic, and phonetic elements. These manuscripts reflect the fusion of linguistic and symbolic representation in healing traditions. Moreover, the evolution of Dongba hieroglyphs as a transitional system from pictographic to phonetic writing is examined by Zhou [15], highlighting the structural variations and writing conventions that define its uniqueness.

Efforts to modernize the script and integrate it into contemporary linguistic systems continue to be explored. Poupard [16] reviews past and ongoing attempts at Unicode encoding, discussing the challenges and prospects of adapting the Dongba script for everyday written communication. These initiatives aim to increase awareness and facilitate the script’s use beyond religious and ceremonial contexts, potentially transforming it into a functional vernacular writing system.

Dongba pictographs are an essential part of the traditional culture of the Naxi people, carrying rich historical and cultural information. Researching Dongba pictographs helps improve the efficiency of interpreting Dongba characters and has significant academic value for cultural, linguistic, and interdisciplinary research. Traditional Dongba pictograph recognition mainly includes three steps: image processing, feature extraction, and classification. Image processing includes grayscale, filtering and denoising, binarization, etc.; feature extraction mostly uses the grid and projection methods; classifier models include support vector machine, random forest, etc. With the development of computer image processing and artificial intelligence technology, more and more scholars have begun to try to use deep learning methods to identify Dongba pictographs.

The Dongba script is one of the few surviving pictographic writing systems used by the Naxi people. Previous computational approaches to Dongba pictograph recognition primarily employed traditional feature extraction and classification methods (e.g., Support Vector Machine (SVM), projection methods, and grid decomposition), with recognition accuracies typically ranging from 84.4% to 95.8% [17,18,19,20,21,22,23]. However, these methods struggle with the variability, complexity, and deformation inherent in pictographic characters. Although deep learning methods, such as CNN, have recently emerged, existing research has been limited by the use of small datasets and minimal data augmentation. Therefore, there is a pressing need for robust and scalable solutions to improve accuracy and generalizability. To address these limitations, this study proposes an optimized CNN model (ResNet) trained on a significantly larger dataset (70,000 pictographs) that has been enhanced through advanced data augmentation strategies [24]. Our approach achieves a classification accuracy of 99.43%, considerably surpassing existing methods and providing a powerful tool for cultural heritage preservation and computational linguistics.

Dongba’s pictographic characters reflect the characteristics of agricultural culture and nomadic culture. They contain many pictographic, ideographic, and indicative characters, and their shapes are complex and varied. To improve the recognition rate of characters and prepare for the recognition and research of text semantics, it is necessary to extract similar features and establish classification standards. To this end, this paper first divides Dongba characters into 18 categories based on the shape features of the characters. It uses the CNN algorithm to extract features and automatically classify Dongba pictographic characters. The specific research includes (1) preprocessing the collected text images by image segmentation, normalization, grayscale, filtering and denoising, and binarization; (2) dividing the text images into 18 categories based on the shape features of the characters and manually annotating them; and (3) using the convolutional neural network to extract image features, train the classification model, and test it with test samples.

2. Materials and Methods

This section outlines the methodologies employed in the study, including dataset collection, preprocessing techniques, and the deep learning framework used for Dongba pictograph classification. The research begins with acquiring 70,000 manually annotated Dongba pictograph images, categorized into 18 classes based on shape characteristics. Image preprocessing steps—such as grayscale conversion, noise reduction, and binarization—enhance image quality for model training. A CNN is then employed for feature extraction and classification, leveraging data augmentation techniques to improve recognition accuracy. The study evaluates multiple deep learning architectures, comparing their performance to determine the most effective model for Dongba pictograph recognition. Although the dataset includes 70,000 Dongba pictograph samples, significant inherent variability and complexity remain challenges for recognition tasks. Thus, data augmentation was employed to simulate naturally occurring variations—such as rotation, affine transformations, scaling, and translation—that commonly occur in both printed and handwritten pictographs. This strategy enhances model robustness and generalization, thereby reducing overfitting and ensuring reliable recognition across diverse real-world scenarios. The following subsections detail each methodology phase, ensuring reproducibility and transparency in the classification process.

2.1. Characteristics and Collection of the Dongba Character Dataset

Yu [25] discussed several issues in compiling the “Naxi Dongba Dictionary”. In the document, Dongba hieroglyphs were divided into 18 categories according to the shape characteristics of the characters, including celestial phenomena, geography, plants, birds, beasts, insects and fish, personal names, human affairs, body shapes, clothing, food, residence, utensils, behavior, shapes, numbers, religions, and legendary names. The collected Dongba hieroglyphs were enhanced and manually annotated, and finally, 70,000-character image samples were obtained. Categories with associated characteristics and number of samples are summarized in

Table 1. Categories and associated characteristics of the Dongba pictograph dataset.

Category	Characteristic	No. of Samples
1	Celestial phenomenon	4474
2	Geographical	4253
3	Plant genus	4159
4	Birds	3981
5	Beast	3874
6	Insects and fish	4128
7	Personal name	4529
8	Personnel	4347
9	Form	2987
10	Clothing	3913
11	Diet	4529
12	Residential	4193
13	Utensils	4789
14	Behavior	2789
15	Shape	1319
16	Several names	3819
17	Religious	4110
18	Legendary name	3807

.

2.2. Methodology and Protocol for the Dongba Script Data Acquisition and Subsection

Dongba pictograph image acquisition involves several key stages, including image processing (preprocessing, edge detection, and contour extraction), feature extraction, and text classification. This classification phase encompasses dataset construction, model training, and model testing. The following sections provide details for each of these processing steps, which are shown in Figure 1.

2.2.1. Dongba Pictograph Image Collection and Data Enhancement

The dataset utilized in this study was compiled from three primary sources. First, 669 pages of scanned images were extracted from the Naxi Pictograph Dictionary compiled by Fang [26]. Second, high-resolution photographs of handwritten Dongba characters were obtained through museum collections and field documentation. Third, manual annotation and classification were conducted following the 18-category system outlined by Yu [25]. These combined sources provided a comprehensive and representative dataset suitable for training and evaluating the proposed models. Additionally, the study incorporated original image data of handwritten Dongba pictographs collected through diverse channels, including manuscript scans, high-definition photographs, and images captured at tourist attractions. Throughout the data collection process, particular emphasis was placed on ensuring sample diversity and clarity. Special attention was given to capturing various writing methods of the same pictographic character across different media, thereby enhancing the dataset’s coverage of stylistic variations and deformation patterns.

To address the challenge of limited Dongba pictographic script samples, this study employed a multi-level data enhancement strategy to construct an augmented dataset comprising 70,000 samples. The data enhancement framework is illustrated in Figure 2. The enhancement strategy is designed based on two core principles: (1) preserving the topological integrity of the script and (2) simulating the natural variations observed in actual handwriting.

Before determining the most suitable enhancement technology, this study conducted a comprehensive literature review to evaluate various existing approaches. Then, a comparative analysis was performed to assess their effectiveness, advantages, and limitations. The findings from this analysis are summarized and presented in

Table 2. Comparison of literature on image enhancement methods.

Research Methods	This Program	Guo et al. [17]	Yang and Kang [12]
Enhanced type diversity	9 types	8 types	6 types
Cultural fit verification	√	✗	✗
Improvement in accuracy	+24.5%	+19.7%	+18.2%

Note: ✓ indicates ‘Conform’; ✗ indicates ‘No match’.

, providing a clear basis for selecting the optimal enhancement technology.

The final enhancement pipeline was developed through experimental validation by integrating multiple enhancement techniques. The adopted pipeline consists of the following operations, with corresponding parameters detailed in Table 3.

AugmentationPipeline = [

GeometricTransformations(), # Applies geometric transformations

PhotometricAdjustments(),    # Adjusts photometric properties

NoiseInjection(),                                 # Introduces noise to the data

OcclusionSimulation()                       # Simulates occlusion

]

This structured approach ensures comprehensive data augmentation for improved model robustness.

Table 3. Image enhancement parameter configuration table.

Research Methods	This Program	Guo et al. [17]	Yang and Kang [12]
Geometric transformation	Randomly rotate 90°	N ∈ {0, 1, 2, 3}	50%
	Flip horizontal	-	40%
	Elastic deformation	α = 0.8, σ = 40	30%
Brightness Adjustment	Brightness and contrast adjustment	ΔBrightness ± 20%, ΔContrast ± 20%	70%
	CLAHE histogram equalization	Grid size = 8 × 8	30%
	RGB offset	Δ channel value ± 15	50%
Noise interference	Gaussian noise	σ ∈ [10, 30]	40%
	Compression artifacts	Quality factor ∈ [50, 90]	30%
Occlusion simulation	Random block dropping	Maximum number of holes = 5, size = 24	40%

Table 3. Image enhancement parameter configuration table.

Research Methods	This Program	Guo et al. [17]	Yang and Kang [12]
Geometric transformation	Randomly rotate 90°	N ∈ {0, 1, 2, 3}	50%
	Flip horizontal	-	40%
	Elastic deformation	α = 0.8, σ = 40	30%
Brightness Adjustment	Brightness and contrast adjustment	ΔBrightness ± 20%, ΔContrast ± 20%	70%
	CLAHE histogram equalization	Grid size = 8 × 8	30%
	RGB offset	Δ channel value ± 15	50%
Noise interference	Gaussian noise	σ ∈ [10, 30]	40%
	Compression artifacts	Quality factor ∈ [50, 90]	30%
Occlusion simulation	Random block dropping	Maximum number of holes = 5, size = 24	40%

• Random Rotation (90°): Image rotation randomly selected from angles {0°, 90°, 180°, 270°}.
• CLAHE Histogram Equalization: Contrast limited adaptive histogram equalization applied with an 8 × 8 grid size.
• Elastic Deformation: Image deformation applied using parameter settings α = 0.8, σ = 40.

Figure 3 illustrates the enhancement effects of the typical Dongba pictograph “bird head”. The figure presents (a) the original sample, (b) rotation combined with brightness adjustment, (c) elastic deformation with added noise, (d) occlusion simulation, and (e) color shift with blur processing.

The applied techniques were carefully selected to enhance model robustness in recognizing Dongba pictographs. Rotation and affine transformations were employed to account for the positional variations commonly observed in manuscripts and inscriptions, thus improving the model’s spatial invariance. Scaling and translation techniques were introduced to address inconsistencies in size and alignment typically present in handwritten and scanned images, enabling the model to adapt to diverse character proportions and placements. Elastic deformation was employed to simulate the natural distortions observed in brushstrokes and handwritten text, introducing controlled warping that enhances the model’s tolerance to irregular shapes. Brightness and contrast adjustments, along with Contrast Limited Adaptive Histogram Equalization (CLAHE), were applied to simulate variations in lighting conditions during scanning or photography, thereby enhancing robustness under diverse illumination settings.

Additionally, noise injection and the simulation of compression artifacts were incorporated to replicate image degradation from old manuscripts, low-resolution scans, or file compression, ensuring that the model maintains performance under noisy conditions. Occlusion simulation, such as block dropping, was implemented to prepare the model for recognizing characters with missing parts or smudges, a frequent issue in degraded Dongba documents. The effectiveness of these enhancement strategies was validated through ablation experiments, as presented in

Table 4. Image enhancement validation test.

Research Methods	Accuracy (%)	Loss Value (Guo et al. [17])
Original (7000 samples)	78.20	0.451
Enhanced by (70,000 samples)	92.70	0.919
W/O occlusion	89.70	0.883
W/O photometric	85.10	0.832

. The dataset used for evaluation comprised 7000 original image samples and 70,000 extended samples. Notably, the condition labeled “W/O Occlusion” refers to the exclusion of occlusion simulations, while “W/O Photometric” indicates the absence of photometric tests.

2.2.2. Image Preprocessing

Image preprocessing is a key step prior to the training of the image classification model, which aims to improve image quality, reduce noise interference, and extract features useful for the classification task. The processing workflow starting from collected color images and ending with black-and-white variations is presented in Figure 4. This includes (1) importing a Dongba pictograph image; (2) performing grayscale processing and position normalization on the collected image; (3) using Gaussian filtering to denoise the normalized image; and (4) binarizing the denoised image into a black-and-white image. Figure 5 shows a set of collected samples and their preprocessed variations. With these, the text features become more prominent in black-and-white images than in the originals.

Graying and position normalization of the collected image involves converting it to grayscale and binarizing it with a threshold set at 127. Next, external contour retrieval is performed using the RETR_EXTERNAL parameter in OpenCV (cv2). Once the contour is retrieved, contour approximation is applied to reduce redundant points in the horizontal, vertical, and diagonal directions, retaining only the endpoints using the CHAIN_APPROX_SIMPLE parameter in cv2. After processing the contour, the image matrix is translated by calculating the necessary shift to align the centroid with the center of the image. The translation matrix MMM is constructed based on the differences (dxdxdx and dydydy) between the centroid and the image center, obtained by subtracting the centroid coordinates from the image center coordinates. Finally, the image is translated accordingly, ensuring that the Dongba pictograph is centered in each image. The relevant parameter settings are presented in

Table 5. Parameters for image grayscale conversion and normalization.

Parameter Name	Value Exploited	Description
Binarization threshold	127	Threshold value for converting grayscale image to binary image
Profile retrieval mode	cv2.RETR_EXTERNAL	Retrieving the external contour
Contour approximation method	cv2.CHAIN_APPROX_SIMPLE	Contour approximation method: compress redundant points and keep only endpoints
Target center of mass X coordinate (standard)	image.shape [1]//2	The X coordinate position to which the centroid is aligned, that is, the X coordinate of the image center
Target center of mass Y coordinate (standard)	image.shape [0]//2	The Y coordinate position to which the centroid is aligned, that is, the Y coordinate of the image center
Binarization threshold	127	Threshold value for converting grayscale image to binary image
Profile retrieval mode	cv2.RETR_EXTERNAL	Retrieving the external contour

.

Besides, this image preprocessing workflow continues with an application of adaptive filter in the Gaussian algorithm: Gaussian kernel is set to (5, 5) to limit the filtering range, the kernel standard deviation is initially 0 and then automatically updated. By applying the Gaussian function to each pixel in the Dongba pictograph image and revising its value based on the weighted average of those of its neighbors, noises and outliers can be effectively removed, making the image look softer and smoother. This can distinguish the foreground and background of the Dongba pictograph in an image and highlight the text part, which is conducive to the recognition of Dongba pictographs. This smoothing effect is not limited to the whole but can also perform fine processing on the local area while retaining the edge information. This feature gives Gaussian filtering significant advantages in image noise reduction and blur processing. This study uses three algorithms, Gaussian blur, median blur, and average blur, to intuitively compare the same Dongba pictograph image, as shown in Figure 6.

2.2.3. Edge Detection and Contour Extraction

After those preprocessing steps, the image data can be further analyzed using edge detection and contour extraction techniques. This article uses the Canny edge detection algorithm [27]. The threshold of the Canny edge detection (the threshold suitable for Dongba pictographs is 100 and 200) needs to be fine-tuned according to the specific category of pictographs. Then, the fine contours function is exploited to find the contours, and the Dongba pictographs, after edge detection and contour extraction, are drawn on the original image. The image sample is shown in Figure 7.

2.2.4. Morphological Processing

Some Dongba hieroglyphic images may be broken, damaged, or degraded due to writing adjustments and preservation techniques. As such, morphological processing can help to enhance and restore those images. This study separated and combined different parts of Dongba hieroglyphic images through morphological operations such as dilation, corrosion, and closing operations (i.e., filling small holes, connecting adjacent objects, and smoothing edges). Specific to the current research, image restoration and super-resolution reconstruction [28] were used to restore the original information of Dongba hieroglyphs, thereby improving the clarity and quality of the image.

This process can be summarized as follows. Firstly, in the dilation operation of Dongba pictograph images, a structural element (kernel) is used to dilate the pictograph image. The foreground becomes larger (the foreground is the pictographic part) and then connects and fills the tiny black holes to complete the broken stroke connection repair operation. The structural element is determined by the kernel_size of a 2 × 2 matrix. Then, in the Dongba hieroglyph image erosion operation, the same structural element (kernel) is used to perform an erosion operation on the expanded Dongba hieroglyph image to reduce the white area, which helps to remove small white noise and smooth the boundary of the Dongba hieroglyphs. Finally, the closed operation is used on the Dongba hieroglyphic image. The closed operation mainly fills small holes, connects adjacent objects, and smoothes boundaries. The closed operation can effectively fill the small holes and long black cracks inside the Dongba hieroglyphic image and has a good effect on broken strokes in connected hieroglyphics. For Dongba hieroglyphic strokes that are close to each other, the closed operation can help connect them. Through the combination of dilation and erosion, the outer boundary of the Dongba hieroglyphs can be smoothed, and the noise can be reduced.

2.2.5. Feature Extraction and Classification Process

With the goal of improving the classification accuracy, this paper makes use of deep learning methods to facilitate feature extraction and text classification of Dongba hieroglyphs. CNN is a deep learning model, a feedforward neural network with a deep structure, which can automatically learn discriminative features from a large image dataset [29,30]. Useful features can be identified in Dongba hieroglyphs by repeatedly performing convolution and pooling between the input layer (data reading) and the output layer (classification prediction) and finally performing complete connection processing, thereby realizing target detection and efficient image classification.

(1) Establishment and division of model training dataset

A total of 70,000 samples of Dongba pictographs were collected, processed, and enhanced, and the dataset covers 18 attributes. The dataset randomly divides the model training dataset into training set, validation set, and test set in a ratio of 7:2:1. The complete dataset of 70,000 samples was randomly divided into three subsets: 70% for training, 20% for validation, and 10% for testing. This partitioning strategy aligns with standard deep learning practices, ensuring that the model has sufficient data to learn effectively while also allowing for hyperparameter tuning and unbiased performance evaluation.

(2) Feature extraction and classification training

The feature extraction and classification training of Dongba pictographs are shown in Figure 8 and Figure 9. A CNN slides the convolution kernel in the convolution layer on the input image to perform convolution operations, thereby extracting local features in the image. The convolution operation can be expressed as follows:

s(i, j) = ∑ₘ∑ₙ I(i − m, j − n)K(m,n)

(1)

where

• s(i, j): the output of convolution at position (i, j).
• I(i − m, j − n): the input image pixel intensity at coordinates offset by (m,n).
• K(m,n): the convolution kernel at position (m,n).

This is the discrete convolution formula, commonly used in image processing and signal processing. It calculates how the shape of one function (K) modifies another function (I) at each point (i, j). Among them, I is the input 2D image data, and K is the 2D convolution kernel size.

The pooling layer further reduces the size of the feature map through sampling operations, reduces computational complexity, and effectively reduces overfitting. At the same time, the pooling layer can also retain key information in the feature map and remove redundant data. The maximum pooling used in this article can be expressed as follows:

yᵢⱼ = max(Xₘₙ) m∈Rᵢⱼ

(2)

where

• y_ij represents the value of the pooled feature at location (i, j) in the resulting output matrix after the pooling step.
• X_mn refers to the input feature values at positions (m, n) within a specific region of the feature map.
• R_ij denotes the pooling region or “window” in the input feature map, associated with position (i, j). Typically, this pooling window is a small square (for example, 2 × 2 or 3 × 3).
• max(X_mn) indicates selecting the maximum value from all values within the region R_ij.

y at position (i, j) equals the maximum value of X over all indices (m, n), where m and n are elements within the region Rᵢⱼ.

After being processed by multiple convolutional layers and pooling layers, the feature map is passed to the fully connected layer. The fully connected layer maps the extracted features to the final output category to complete the classification task. The operation of the fully connected layer can be expressed as follows:

z = Wx + b

(3)

where

• z is the output.
• W represents the weights.
• x represents the input.
• b represents the bias term.

Among them, x is the output feature vector of the previous layer, W is the weight matrix, b is the bias term, and z is the output of the fully connected layer.

Figure 8. Process of text feature extraction and classification.

Figure 8. Process of text feature extraction and classification.

Figure 9. Workflow of the model classification process.

Figure 9. Workflow of the model classification process.

(3) Investigated CNN models

Five classic CNNs, LeNet-5 [31], AlexNet [32], VGGNet [33], ResNet [34], and Google Inception Net [35], were selected as classification models. Each model improves its performance through different structural improvements, as shown in

Table 6. Specifications and characteristics of CNN models used in the experiment.

CNN Model	Total Layers	Key Characteristics	Applicability
LeNet-5	7 layers (2 convolutional, 2 pooling, 3 fully connected)	Basic text feature extraction	Good for smaller datasets
AlexNet	8 layers (5 convolutional, 3 fully connected)	ReLU activation, dropout, fast training	Effective for complex features
VGGNet	19 layers (multiple 3 × 3 convolutional, max-pooling layers)	Deep network for detailed features	Captures detailed pictograph nuances
ResNet	Deep layers with residual blocks	Addresses gradient vanishing and degradation	Best performance for deep networks
Google Inception Net	22 layers (multi-scale convolutional kernels)	Multi-scale feature extraction	Good for varied-size pictographs

.

The reason for choosing these five CNN models for the experiment is that they each have unique advantages and applicable scenarios. LeNet-5, as a classic convolutional neural network, is suitable for processing smaller image datasets; AlexNet and VGGNet have deeper network structures and can capture complex image features; ResNet solves the training problem of deep networks through residual learning mechanisms and has better stability and accuracy; and Google Inception Net improves the generalization ability of the model through multi-scale feature fusion. In addition, these models can overcome the shortcomings of traditional algorithms. Traditional algorithms, such as grid and projection methods, often have difficulty capturing characters’ complex features and deformation when processing Dongba pictograms. The CNN models can improve recognition accuracy by automatically learning features and performing advanced processing.

(4) Training Configuration

When different models are used to recognize Dongba images, the parameters adopted are slightly different.

Table 7. Hyperparameter configuration for different model trainings.

Model	Input Image Size	Sample Batch Size	Maximum Training Cycle	Learning Rate	Optimizer
LeNet-5				0.0005	Adam
AlexNet				0.001	Adam
VGGNet	64 × 64	512	100	0.0001	SGD
ResNet				0.0005	Adam
Google Inception Net				0.001	RMSprop

lists the hyperparameters used in training each model. Specifically, the initial learning rate was set to 0.001 and decayed using a step learning rate schedule with a decay factor of 0.1 every 25 epochs. We applied L2 regularization (weight decay) with a coefficient of 1 × 10⁻⁴ to reduce overfitting and improve generalization. The optimizer used was Adam, with β₁ = 0.9 and β₂ = 0.999. Additionally, we implemented dropout with a rate of 0.5 in the fully connected layers to further mitigate overfitting. The training was conducted for 75 epochs with a batch size of 512, as previously reported, and early stopping was employed with a patience threshold of 10 epochs based on validation loss.

3. Results

This section presents the experimental results of the proposed CNN-based classification method for Dongba pictographs. The performance of various deep learning models, including LeNet-5, AlexNet, VGGNet, ResNet, and InceptionNet, is evaluated using a dataset of 70,000 pictograph samples categorized into 18 classes. Key performance metrics such as classification accuracy, loss function values, and computational efficiency are analyzed. Additionally, the impact of data augmentation techniques on model robustness is examined. Comparative analysis highlights the superiority of the ResNet model, which achieves the highest accuracy while maintaining computational efficiency. The findings demonstrate the effectiveness of deep learning in enhancing Dongba pictograph recognition, supporting its application in cultural heritage preservation and computational linguistics.

3.1. Model Classification Training and Optimization

Due to limited experimental equipment, large datasets of typical models in CNN models, and slow training, 10 rounds were conducted for 18 fixed types of pictograms. The data in each round were the data listed in Table 1.

Table 8. Performance comparison of various models in Dongba pictograph recognition.

Model	Accuracy (%)	Loss Value	Training Time (Hours)
LeNet-5	20.66	2.357	1.5
AlexNet	97.72	0.125	6.0
VGGNet	97.72	0.094	8.0
Google Inception Net	68.75	0.431	7.5
ResNet	99.43 (baseline)/99.84 (peak)	0.038	7.5

presents a comparative analysis of the classification performance of various CNN models applied to Dongba pictographic images, highlighting their highest achieved accuracy and corresponding loss values. It also includes a total training time (in hours) column for each model under identical training conditions. This column facilitates a more comprehensive assessment of the trade-offs between model performance, as reflected by accuracy and loss, and computational efficiency. All training times were recorded using the same hardware configuration to ensure consistency and comparability. The inclusion of this information enhances the interpretability of the experimental outcomes and further substantiates the selection of ResNet as the most effective model, striking a balance between accuracy and training efficiency. It is important to note that both hardware and software settings for this experiment are as follows:

Hardware environment: Intel Core i7-10700 CPU @ 2.90 GHz, 20 GB memory, NVIDIA GeForce GTX 1050Ti GPU (4 GB video memory) (Intel, Santa Clara, CA, USA).

Software environment: Windows 10 operating system, Python 3.9.7 development environment. Image processing operations are implemented using OpenCV-Python 4.5.5.64, and the deep learning framework is implemented using PyTorch 1.10.2.

Under the premise that the hyperparameters were relatively fixed, the same batch of samples was trained multiple times using different models.

Table 9. Comparative training performance of various CNN models.

Model	Accuracy	Loss Value	Training Time (One Hour)
LeNet-5	20.66%	1.9681	1.5
AlexNet	97.72%	0.1286	6.0
VGGNet	97.72%	0.0998	8.0
ResNet	99.43%	0.2040	7.5
Google Inception Net	68.75%	0.9576	9.0

lists the performance of other models in the Dongba hieroglyph classification task.

As can be seen from Table 9, the ResNet model performs best in the Dongba pictograph classification task, with an accuracy of 99.43% and a relatively short training time. AlexNet and VGGNet also have high accuracy but take longer to train. LeNet-5 has a low accuracy rate because of its shallow network structure and its inability to fully capture character features. Although Inception Net has a complex structure and high computational cost, its performance in this task is not ideal, which may be related to the fact that its multi-scale feature fusion method is not suitable for the characteristics of Dongba pictographs.

Comparing the training conditions and results of each model in the convolutional neural network model, it is not difficult to find that for Dongba pictograph images, the ResNet model, AlexNet model, and VGGNet model have achieved high accuracy, and their loss function values are small, which can obtain good classification results. Among them, the ResNet model performs best. However, for the classification of Dongba pictograph images, it is not easy to provide many training samples for these better models. Even if a relatively large training dataset can be provided, it will be limited by the computing power required by the model itself. Therefore, it is necessary to make up for the shortcomings of the typical convolutional neural network model in the specific scenario of Dongba pictograph image classification and then obtain better training and classification results under relatively limited datasets. Based on this, this paper continuously adjusts the structure and parameters based on the ResNet model. It uses a greedy algorithm to solve the approximate optimal solution of the loss function.

For the ResNet model, the network structure and parameters were adjusted to optimize the model performance further. Specifically, the number of residual blocks was increased to make the network deeper; simultaneously, the size and number of convolution kernels were adjusted to capture the characteristics of Dongba pictographs better. In addition, a greedy algorithm was used to solve the approximate optimal solution of the loss function. The greedy algorithm gradually optimizes the parameters of each residual block so that the model gradually converges to the optimal solution during the training process.

To prevent the model from overfitting, the L2 norm [36] of the weight is added as a penalty term in the loss function. L2 regularization makes the model smoother by penalizing larger weight values, thereby improving the generalization ability. In the specific implementation, the L2 regularization term is combined with the cross-entropy loss function to form a new loss function for optimization. Adding L2 regularization to the loss function prevents the model from overfitting, and the model’s generalization ability is significantly improved. The experimental results show that L2 regularization reduces the overfitting phenomenon of the model on the training set and enhances the model’s performance on the validation and test sets. This enables the model to better adapt to different writing styles and Dongba hieroglyphs with large deformations and has more substantial practical application value.

Specifically, while VGGNet and AlexNet achieved high accuracy (97.72%), they required longer training durations (8.0 and 6.0 h, respectively) and exhibited less stability in handling structural variation. LeNet-5, although computationally efficient, suffered from significantly lower accuracy (20.66%) due to its shallow architecture, which limited its ability to capture the complex features of Dongba pictographs. InceptionNet, despite its multi-scale feature extraction design, performed suboptimally in our setting (68.75%), likely due to the model’s sensitivity to input scale inconsistency.

In contrast, ResNet not only outperformed all models in classification accuracy (baseline 99.43%, peak 99.84%) but also demonstrated a favorable balance between depth and efficiency, enabled by its residual learning mechanism. This architecture effectively mitigates vanishing gradient issues in deep networks, supporting stable training even with complex and variably deformed pictographs. Additionally, ResNet achieved high performance with a moderate training time of 7.5 h, offering a practical compromise between computational demand and classification accuracy. These factors collectively justify our selection of ResNet as the most suitable architecture for recognizing Dongba pictographs.

3.2. Model Classification Training Results and Analysis

The experimental results indicate that the ResNet model achieved the highest performance in the Dongba pictograph classification task. Specifically, extensive experiments were conducted using various parameter combinations to optimize the training model. Ultimately, by fixing other parameters and varying the number of epochs and batch size, the approximate optimal solution of the loss function was obtained when the number of epochs was set to 75 and the batch size to 512. Under these conditions, the classification accuracy reached 99.84%, as illustrated in Figure 10. The proposed model achieved a baseline classification accuracy of 99.43%, while its performance peaked at 99.84% under optimized training conditions—specifically, with 75 training epochs and a batch size of 512.

In contrast, LeNet-5 and InceptionNet exhibited suboptimal performance, achieving only 20.66% and 68.75% accuracy rates, respectively. In the classification and recognition of Dongba pictograms, ResNet effectively mitigated the gradient vanishing problem through residual connections, allowing for the training of a deeper network capable of capturing intricate features of Dongba characters. Consequently, ResNet significantly outperformed models such as AlexNet and VGGNet. The test confidence support for these findings is presented in Figure 11 and Figure 12.

Simultaneously, the model’s performance was further optimized by refining the network architecture and hyperparameters, incorporating regularization techniques, and employing greedy algorithms to approximate the optimal solution of the loss function. While this study achieved high classification accuracy, certain limitations remain. First, despite the large dataset, it primarily consists of printed Dongba pictographs, with a relatively small number of handwritten samples. This imbalance constrains the model’s ability to recognize handwritten characters accurately. Second, the model’s capacity to differentiate similar characters requires further enhancement, mainly when the character shapes exhibit high similarity. Given that Dongba pictographs are inherently hand-painted images with substantial visual resemblance, the model tends to overfit after fewer training epochs, as illustrated in Figure 13.

Data annotation is conducted incrementally, with the sample size gradually increasing. As the dataset expands, the model’s test results become progressively more reliable, as Figure 14, Figure 15, Figure 16 and Figure 17 illustrate.

For hand-painted Dongba pictographs, the model reaches a point where it can no longer learn additional invariant features. To address this, while augmenting the dataset with more existing hand-painted Dongba pictographs, regularization is applied to the Dongba pictograph image convolutional neural network model. Specifically, the L2 norm of the weights is introduced as a penalty term in the loss function to prevent excessive weight magnitudes, thereby reducing model complexity. By incorporating this penalty term while maintaining a high recall rate, the model is evaluated using 700 newly collected hand-painted Dongba pictographs, as illustrated in Figure 18. The confusion matrix reveals that the sky and square categories are more frequently misclassified. This misclassification is attributed to the nature of Dongba pictographs, which often employ geometric shapes, such as circles and squares, to represent these concepts, leading to increased ambiguity between the two categories.

The rounds with low recognition accuracy were identified and analyzed. By examining the recognition process log, as illustrated in Figure 19, it was observed that recognition confidence was consistently low. Several key factors contributed to this issue: (1) significant character deformation, making it difficult for the model to recognize characters accurately; (2) high similarity between characters, leading to confusion, especially when the handwriting lacks standardization; and (3) poor image quality, which hinders accurate feature extraction.

The analysis reveals that these misclassifications are primarily due to visual similarity in stroke structure and spatial arrangement between certain pictographs. For instance, “sky” and “square” both share rectangular outlines and internal symmetry, which may lead to overlapping activation patterns in intermediate layers of the CNN. Additionally, some errors can be attributed to intra-class variability—such as handwritten distortions, partial occlusions, or style-specific glyph variations—that reduce the discriminative power of feature representations.

To support this analysis, representative misclassified image pairs have been included in Figure 19, highlighting the visual ambiguity contributing to model confusion. We have also discussed possible strategies to mitigate such errors, including hierarchical classification, attention-based visual refinement, or domain-adaptive training on stylistic variants of Dongba characters.

Several improvement strategies can be implemented to mitigate the aforementioned challenges. First, enhancing the diversity of training samples can ensure broader coverage of various deformation scenarios. Second, incorporating advanced techniques, such as attention mechanisms, can enhance the model’s capability to differentiate visually similar characters. Third, leveraging image enhancement techniques, including image restoration, can improve overall image quality and facilitate more accurate recognition.

Table 10. Benchmarking of methodological differences.

Study	Model Used	Dataset Size	Data Augmentation	Accuracy (%)
Kadam et al. (2020) [37]	Five different CNN architectures	MNIST (60,000 training images and 10,000 testing images) and Fashion-MNIST (60,000 training images and 10,000 testing images)	Activation functions, optimizers, learning rates, dropout rates, and batch sizes.	93.56
Chen and Zhu [39]	Improved ResNet/Baseline Models	CIFAR-10 (60,000 images)/CIFAR-100 (60,000 images)	Images were bilinearly upsampled to 40 × 40, cropped to 32 × 32, and augmented solely through random rotation.	95.11 (CIFAR-10)/76.37 (CIFAR-100)
Liang [40]	ResNet-20/Inception-ResNet v1	CIFAR-10 dataset with 60,000 32 × 32 RGB color images	ImageDataGenerator	88.28
Zhang [22]	CNN (custom)	~2500	Not specified	95.80
Xie and Dong [23]	ResNet (20-layer)	536 (tested on 94)	None	93.58
Luo et al. [24]	Improved ResNet	Not specified	Basic (rotation)	98.65
Isohanni [38]	ResNet-34	7855 RGB images	Rotations (90, 180, and 270 degrees), vertical and horizontal flips, and a 0.2 zoom option	99.28
This Study	Optimized ResNet	70,000	Comprehensive (rotation, scaling, occlusion, photometric, etc.)	99.43 (baseline)/99.84(peak)

presents a benchmarking analysis of diverse CNN-based image classification studies, highlighting variations in model architectures, dataset sizes, data augmentation techniques, and resulting accuracies. Kadam et al. [37] achieved 93.56% accuracy using five CNN architectures on the MNIST and Fashion-MNIST datasets, emphasizing the importance of parameter tuning over data augmentation. Zhang [22] attained 95.80% accuracy with a custom CNN on a relatively small dataset (~2500 images) despite unspecified augmentation. In contrast, Xie and Dong [23] utilized a 20-layer ResNet on only 536 images without augmentation, achieving an accuracy of 93.58%. Luo et al. [24] achieved a performance improvement to 98.65% using a modified ResNet with basic rotation-based augmentation. Isohanni [38] further improved accuracy to 99.28% with ResNet-34 and a diverse augmentation strategy that included geometric and zoom transformations. Chen and Zhu [39] applied improved ResNet models to CIFAR datasets, reaching 95.11% on CIFAR-10 and 76.37% on CIFAR-100 with minimal augmentation. Liang [40] reported 88.28% accuracy using ResNet-20 and Inception-ResNet v1 with automated augmentation via ImageDataGenerator. Notably, the present study surpasses all prior efforts, achieving 99.43% baseline and 99.84% peak accuracy with an optimized ResNet trained on 70,000 images and enhanced through comprehensive augmentation strategies encompassing rotation, scaling, occlusion, and photometric variation. These findings underscore the critical role of both dataset scale and diverse augmentation in maximizing the performance of CNNs.

Table 11. Individual impact of data augmentation techniques.

Augmentation Technique	Accuracy (%)
No augmentation	92.70
Rotation only	95.50
Scaling only	95.00
Occlusion simulation only	94.10
Photometric adjustment	93.80
Elastic deformation	94.30
Full augmentation pipeline	99.43

presents the results of a comparative analysis in which each data augmentation technique was individually applied to the training dataset, with all other parameters held constant. The experimental findings indicate that rotation and scaling had the most pronounced effects on improving model generalization, yielding accuracy gains of approximately 2.8% and 2.3%, respectively. While occlusion simulation and photometric adjustment produced more modest improvements, they contributed meaningfully to the model’s robustness, particularly in handling image quality variations and partial character degradation. These results underscore the critical role of a comprehensive augmentation pipeline in enhancing model performance and achieving the final baseline accuracy of 99.43%.

4. Discussion

This study presents a convolutional neural network (CNN)-based approach for the classification of Dongba pictographs, achieving a classification accuracy of 99.43%. The model was trained on a dataset of 70,000 manually categorized samples, divided into 18 distinct classes based on shape characteristics. The findings demonstrate that CNN-based methods, particularly ResNet, outperform traditional classification techniques such as Support Vector Machines (SVM) and other deep learning models, including LeNet-5, AlexNet, VGGNet, and InceptionNet. By incorporating data augmentation techniques (e.g., rotation, affine transformation, scaling, and translation), the study enhances the model’s robustness and generalization capability.

The experimental findings demonstrate that the optimized ResNet model achieved superior performance, effectively addressing challenges such as gradient vanishing and overfitting through residual learning mechanisms. When benchmarked against previous studies—such as Kadam et al. [37], Chen and Zhu [39], Liang [40], Zhang [22], Xie and Dong [23], Luo et al. [24], and Isohanni [38]—the proposed model attained a notably higher recognition accuracy of 99.43%. This result surpasses the reported accuracies of 93.56%, 95.11% (CIFAR-10)/76.37% (CIFAR-100), 88.28%, 95.80%, 93.58%, 98.65%, and 99.28%, respectively. These improvements underscore the effectiveness of deep learning approaches, particularly optimized residual networks, in accurately recognizing complex pictographic characters characterized by substantial visual variability.

The study aligns with previous research on Dongba pictograph recognition, reinforcing the value of deep learning techniques in improving classification accuracy. Earlier works primarily employed feature extraction and machine learning methods, such as topological structure-based methods [20], grid decomposition algorithms [12,13], and Histogram of Oriented Gradients (HOG) feature extraction with SVM [21]. However, these approaches exhibited limited adaptability to pictographic variations and suffered from challenges related to feature engineering.

More recent studies (e.g., Luo et al. [24]) attempted to improve classification performance using enhanced ResNet architectures, achieving 98.65% accuracy. Our study extends this line of research by utilizing a larger dataset and introducing advanced data augmentation techniques, further enhancing classification accuracy and robustness.

Additionally, this research contributes to ongoing efforts in cultural heritage preservation and computational linguistics [9,16]. The application of CNNs to Dongba pictographs supports broader digital humanities initiatives, providing a scalable and efficient framework for automated script recognition.

Despite the study’s success in Dongba pictograph recognition, several limitations were noted. The dataset predominantly comprises printed Dongba pictographs, which limits the model’s generalizability to handwritten or historical inscriptions. Although some handwritten samples were included, they constitute a small proportion of the overall dataset. The model also struggles to recognize highly similar pictographs due to minor structural variations. This challenge is reflected in the confusion matrix and highlights the visual resemblance among certain Dongba symbols.

Furthermore, the study does not fully account for real-world degradations in historical texts, such as ink smudging, erosion, or incomplete characters, which can affect recognition accuracy in practical applications [41]. Additionally, computational constraints present another challenge. At the same time, deep learning models improve classification accuracy. They require significant computational resources for training and inference, making implementing these models on low-power devices challenging.

The research has significant implications for both academic and practical applications. Regarding cultural heritage preservation, the proposed CNN model aids in digitally archiving and preserving Dongba pictographs, facilitating automated transcription and indexing of historical manuscripts. For computational linguistics and NLP, future advances could integrate pictographic recognition with semantic interpretation, enabling a deeper understanding of the grammatological and phonological aspects of the Dongba script. The study validates the use of CNNs for low-resource languages and ancient scripts, paving the way for their application in other endangered writing systems, such as Oracle Bone Script and Mayan Hieroglyphs. Enhanced OCR systems could also benefit from incorporating attention mechanisms and hybrid deep learning models, like Transformers, to improve pictograph recognition, especially in noisy and degraded texts.

The study’s introduction highlighted the challenges of Dongba pictograph recognition and the limitations of traditional feature extraction methods and classifier models. The study’s findings address these challenges by demonstrating that deep learning methods, especially ResNet, significantly improve recognition accuracy. The study builds on previous research by expanding the dataset, introducing data augmentation, and exploring broader implications for cultural preservation and AI-driven linguistic research. The potential of deep learning in enhancing Dongba pictograph classification is underscored by achieving an unprecedented accuracy of 99.43%, with the ResNet model outperforming other CNN architectures and traditional machine learning techniques. The findings contribute to computational linguistics, cultural heritage preservation, and AI-based pictograph recognition.

The inclusion of LeNet-5 and other foundational CNN architectures, such as AlexNet and VGGNet, provides baseline comparisons against advanced methods, like ResNet, to systematically demonstrate the performance gains from architectural improvements. While YOLO (You Only Look Once), a prominent modern model designed primarily for object detection tasks, was not directly compared due to its primary focus on real-time bounding-box detection, its potential applicability to real-time detection of pictographs presents an interesting avenue for future research.

To advance this research further, future studies should focus on expanding the dataset to include more handwritten and degraded Dongba pictographs, thereby improving real-world recognition capabilities. They should also integrate attention mechanisms like Vision Transformers to enhance feature extraction and classification accuracy. Developing multimodal approaches that combine visual recognition with linguistic analysis could aid in the semantic understanding of the Dongba script. Additionally, applying advanced image restoration techniques could improve the recognition of damaged or historical pictographs. By addressing these challenges, future research can bridge the gap between AI and cultural heritage studies [42], ensuring that the Dongba script and other endangered writing systems are preserved and accessible for future generations.

5. Conclusions

This study demonstrates the effectiveness of a convolutional neural network (CNN)-based approach, with the ResNet architecture notably enhancing the accuracy of Dongba pictograph recognition. The proposed model achieved a baseline classification accuracy of 99.43%, while its performance peaked at 99.84% under optimized training conditions—specifically, with 75 training epochs and a batch size of 512. This research enhances the model’s robustness and generalization by leveraging a large-scale dataset of 70,000 samples and employing data augmentation techniques, outperforming traditional feature extraction methods and previous deep learning models. The study contributes to the field by addressing the limitations of earlier works, filling a critical research gap in pictographic script recognition, and providing a scalable framework for cultural heritage preservation and computational linguistics. Despite its success, limitations remain, including the model’s difficulty in distinguishing highly similar characters, its reliance on predominantly printed samples, and the challenges posed by historical text degradation, all of which present opportunities for future research in expanding datasets, integrating attention mechanisms, and refining image restoration techniques. Ultimately, this research advances artificial intelligence applications in script recognition and supports the broader goal of preserving endangered writing systems, fostering interdisciplinary collaboration between linguistics, deep learning, and digital humanities, and ensuring the accessibility of Dongba pictographs for future generations.

The primary innovations of this study are threefold. Firstly, we created a novel, large-scale dataset of 70,000 Dongba pictograph samples, which is dramatically larger than previously reported datasets, facilitating more profound and generalized learning. Secondly, a tailored, multi-level data augmentation approach was developed and validated specifically for the Dongba script, effectively handling inherent variability and deformation in pictographs. Lastly, the study optimizes and customizes the ResNet CNN architecture through strategic adjustments in residual blocks and optimization algorithms, resulting in an unprecedented classification accuracy of 99.43%. These innovations collectively set a new benchmark for Dongba pictograph recognition, providing robust methodological contributions with clear practical implications for the preservation of cultural heritage and computational linguistics.