Abstract
To address the decline in detection accuracy caused by the degradation of grayscale features under environmental interference, a lightweight detection model driven by grayscale characterization, YOLOv8n-CoalGangue, is proposed based on an in-depth analysis of the dynamic variations exhibited by grayscale features. First, grayscale histograms are used to quantitatively evaluate the effects of illumination changes and moisture conditions on feature distributions, revealing that global grayscale aliasing and local texture degradation are the key visual feature bottlenecks. Guided by these unique findings, targeted technological innovations are integrated into the developed architecture. HGNetV2-G, which incorporates the GhostNet principle, is used as the backbone to reduce the incurred computational cost while preserving the core feature extraction ability of the model. A mixed local channel attention (MLCA) mechanism is introduced in the neck to filter background noise and focus on local high-frequency features, which helps overcome global grayscale aliasing issues. In addition, a DGFPN-based feature fusion network is constructed by combining RepGFPN and DySample, together with lightweight shared convolution detection (LSCD), which compensates for the loss of multiscale grayscale details without increasing the imposed parameter burden. Furthermore, the PIoUv2 loss function improves the bounding-box regression process in dense overlapping scenarios. Experimental results show that the proposed model achieves an mAP@50 of 97.2% with a 32% reduction in the number of parameters required (only 2.1 M). It also demonstrates strong robustness under six extreme industrial conditions, such as low illumination and coal dust occlusion, confirming the effectiveness of the design driven by grayscale characterization for practical green mining applications.
1. Introduction
As a primary global fossil fuel, coal plays a pivotal role in power generation and industrial tasks. However, in standard underground mining scenarios, raw coal typically contains 20–25% gangue [1]. Combusting gangue alongside coal not only significantly reduces the efficiency of thermal energy conversion but also emits greenhouse gases and toxic substances, posing dual threats to the environment and human health [2]. Consequently, the efficient and clean separation of gangue has become a pressing technical bottleneck in the intelligent development of coal mining methods [3,4,5].
The existing coal and gangue separation frameworks rely primarily on manual picking, wet separation [6], and dry separation [7]. While manual picking is a labor-intensive process and wet separation causes water pollution and high resource consumption levels, dry separation has emerged as the preferred eco-friendly alternative. Within this domain, image recognition-based methods have gained tremendous traction. These systems capture visual data via industrial cameras [8], process them through detection networks, and execute separation using robotic manipulators [9], achieving high automation and precision levels.
The YOLO series has been extensively applied in vision-based coal and gangue detection systems [10]. Nevertheless, practical belt conveyor environments present formidable challenges. Factors such as target occlusion due to material stacking, motion blur caused by high-speed conveyances, and severe environmental disturbances severely degrade the performance of the existing detection networks.
In view of the above problems, this study first quantitatively analyzes the dynamic variations in the characteristics of coal and gangue images under interference conditions. On the basis of these findings, an improved model, i.e., YOLOv8n-CoalGangue, is proposed to achieve high-precision detection in complex industrial environments. The main contributions of this paper are as follows.
- Through a quantitative evaluation, the dynamic evolution trends of the grayscale features of coal and gangue are deeply analyzed under interference conditions such as illumination and moisture. The identification of the key visual feature bottlenecks, specifically global grayscale aliasing and local texture degradation, provides crucial grayscale characterization priors for the subsequent directional network optimization procedure.
- Guided by the grayscale characterization findings, a customized lightweight model called YOLOv8n-CoalGangue is proposed. HGNetV2-G is utilized as the backbone to extract core features with an extremely low computational cost. A mixed local channel attention (MLCA) mechanism is integrated to filter background noise and explicitly overcome global grayscale aliasing. A DGFPN is designed to compensate for the loss of multiscale feature details. Finally, a lightweight shared convolution detection (LSCD) module and the powerful intersection over union v2 (PIoUv2) loss function are introduced to optimize the localization accuracy achieved in dense stacking scenarios.
- A dedicated coal and gangue detection dataset is constructed to comprehensively evaluate the proposed model. The experimental results demonstrate that the YOLOv8n-CoalGangue model achieves superior detection accuracy compared with existing methods while significantly reducing the required parameter count and the imposed computational load. Furthermore, the exceptional robustness of the model is systematically verified under various extreme multi-interference conditions.
2. Related Work
In coal preparation plants, a robust method for detecting coal and gangue based on visual signals is essential for ensuring high efficiency and system stability levels [11,12]. Early studies focused mainly on traditional computer vision methods, including contour detection [13], multifractal spectrum analysis [14], and machine learning approaches based on surface grayscale and gloss features [15,16]. Although they are effective in controlled environments, these methods rely heavily on handcrafted features and struggle to maintain high robustness under complex and dynamic industrial conditions.
With the widespread adoption of convolutional neural networks (CNNs), deep learning has become the dominant approach. On the one hand, image segmentation models, such as U-Net, DeepLabV3+, and their variants [17,18,19,20], have been used to extract coal and gangue regions at the pixel level, even under harsh conditions such as heavy dust. However, segmentation tasks are computationally expensive, making it challenging to deploy such methods on edge devices in real time.
Therefore, object detection has become a more practical solution, and recent studies have widely explored improved YOLO-based architectures. For example, for low-light underground environments, researchers have proposed methods using WaveNet [21], multiobject generation models [22], and YOLO models in combination with dehazing preprocessing [23]. More recently, Yang et al. [24] and Gao et al. [25] improved YOLOv5s and YOLOv8 through multiscale fusion and optimized loss functions, achieving impressive outcomes. In addition, considering deployment constraints, Cao et al. [26] reduced the size of their model to 4.76 M by simplifying its feature fusion structure. Beyond visible light, alternative sensing modalities, such as dual-energy X-ray images [27,28], multispectral imaging scans [29,30], and laser speckle techniques [31], have been explored. These methods can mitigate surface feature degradation, but they significantly increase the incurred hardware costs and the complexity of maintenance processes.
Despite these advances, important challenges remain. Many studies have aimed to maximize the attained values by stacking complex attention mechanisms or designing heavy multiscale fusion networks. These approaches are often based on general computer vision techniques and tend to treat neural networks as black boxes. As a result, quantitative analyses of how grayscale features degrade under severe coal dust occlusion are lacking.
To address these issues, the YOLOv8n-CoalGangue framework introduces an optimization strategy guided by grayscale characterization. By combining a customized HGNetV2-G backbone with MLCA and DGFPN modules, the model effectively handles global grayscale aliasing and feature degradation issues caused by environmental factors. Moreover, it achieves strong parameter compression effects. As a result, the model maintains high detection accuracy and a fast inference speed while satisfying the strict edge deployment requirements imposed in mining environments.
3. Materials and Methods
3.1. Dataset Acquisition and Data Augmentation
To obtain a coal and gangue dataset that was more in line with actual working conditions, coal and gangue of different sizes on a belt conveyor were selected as data acquisition objects. During the data acquisition process, by adjusting the relative positions of the coal and gangue, the illumination intensity level and the camera angles, the dataset could include a wider range of working conditions, thereby enhancing its representativeness and universality. The data were acquired by an industrial camera. Factors such as the overlap and occlusion of coal and gangue were considered, and different block sizes and shapes were used to form different combinations. Finally, a coal and gangue dataset was obtained, including a total of 5519 photos and featuring diverse ranges of quantities and scales for both the coal and gangue samples.
In this study, the initial dataset consisting of 5519 samples was first randomly partitioned into training, validation, and test sets at a ratio of 7:1:2 (comprising 3866, 551, and 1102 images, respectively). During the training process of YOLOv8, Mosaic data augmentation technology was applied exclusively to the training set to combine multiple photos into new training samples. The training data was further expanded through rotation, translation, clipping, scaling, and combination operations. To ensure a more comprehensive training process, manual data augmentation was also performed on the training images, including adding partial occlusion, adding noise, adjusting brightness and contrast levels, and applying motion blur. These operations simulated the challenging conditions of industrial environments, such as dust interference and high-speed conveyor movement. Through these specific operations, the training set was expanded to 9385 images, bringing the total number of images in the final dataset to 11,038, as shown in Figure 1. The validation and test sets were kept in their original, unaugmented states. This dataset construction strategy enhances the generalizability of the model.
Figure 1.
Data augmentation methods used in the dataset.
3.2. Grayscale Characterization of Coal and Gangue Under Various Environmental Factors
Owing to the harsh environment involved in the process of coal and gangue detection, illuminators, dust removal nozzles and pulverized coal affect the identification of coal and gangue. During CMOS imaging, the image quality of the same coal or gangue sample is significantly affected by varying light intensities, moisture states, and surface conditions. Under low-light conditions, photon noise and dark current noise become prominent. When the surfaces of coal and gangue are covered with water or pulverized coal, the light undergoes Mie or Rayleigh scattering [32] upon making contact with these surfaces, resulting in overall image blurring or localized halos.
During digital image processing, a two-dimensional image can be defined as a two-dimensional discrete function f(x,y), where x and y represent the spatial coordinates of the image. The grayscale histogram of an image is a core statistical tool in digital image processing tasks, where it can be used to quantitatively analyze the distribution of grayscale levels in an image and is an effective tool for analyzing the characteristics of coal and gangue images acquired under different influencing factors.
Let rk denote the grayscale of an L-level grayscale digital image f(x,y); the nonnormalized histogram of f is defined as follows:
where nk denotes the number of pixels in f with a grayscale of rk, and in the commonly used 8-bit grayscale image setting, L = 256.
Since the sizes of different images are not the same, to facilitate comparison, the normalized histogram of f is defined as follows:
where M and N denote the rows and columns of the input image, respectively. In a normalized histogram, the sum of the p(rk) equals 1, representing a probability distribution that provides a more accurate characterization of the grayscale information of the image. The normalized histogram is suitable for implementing a fast hardware implementation. For different image sizes, the normalized histogram removes the interference caused by the width and height of the image and is more suitable for analyzing the features of coal and gangue images under different influencing factors.
3.3. YOLOv8n-CoalGangue Algorithm for Coal and Gangue Detection
In the YOLO series of networks, the YOLOv8 architecture includes a series of models that are scaled by depth and width factors (n, s, m, l, and x). Among them, YOLOv8n (nano) has the most compact network size, resulting in the smallest number of parameters and the lowest computational cost. In this study, YOLOv8n was explicitly selected as the practically driven baseline model, as intelligent sorting systems place greater emphasis on the edge deployment capabilities of equipment in mining environments, while the high-speed dynamics of belt conveyors impose relatively low requirements. However, although YOLOv8n can satisfy these real-time and lightweight demands, its architecture indicates that its feature extraction ability is highly vulnerable to the grayscale variations caused by environmental disturbances such as insufficient illumination and moisture levels. Therefore, the YOLOv8n-CoalGangue network was developed and proposed specifically for the detection of coal and gangue. To minimize the parameter overhead, a lightweight shared convolutional detection head was employed, and an updated loss function was used to accelerate the convergence rate of the model. As shown in Figure 2, these enhancements improved the detection accuracy of the model while reducing its overall parameter count.
Figure 2.
Network architecture of YOLOv8n-CoalGangue.
3.3.1. HGNetv2-G Backbone Network
In the backbone network part, because coal and gangue images have obvious gray characteristic differences under uniform illumination conditions, the model obviates the need for excessively redundant parameters. Consequently, a feature extraction network based on the HGNetv2 [33] architecture is proposed. To reduce the complexity of the model while maintaining its inference speed, the network is optimized by incorporating GhostConv [34]. This integration scheme results in an enhanced architecture, i.e., HGNetV2-G, as shown in Figure 3, which is designed to facilitate the effective extraction of multiscale coal and gangue features.
Figure 3.
YOLOv8n-CoalGangue HGNetV2-G backbone network structure.
At the input level of the HGNetv2-G backbone, the HGStem module facilitates the rapid extraction of features from the raw data. This is followed by four hierarchical downsampling stages composed of Ghost-HGBlock and depthwise convolution (DWConv) mechanisms. These stages are designed to progressively attenuate the spatial resolution while simultaneously expanding the channel dimensions.
GhostConv adjusts the feature map to half of the original number of output channels through a convolution operation executed in the first step and performs cheap operations and identity mapping to the next level at the same time. By using a concatenation strategy to replace traditional convolution modules, GhostConv effectively reduces the computational overhead and FLOPs of the network. Specifically, given an input tensor X ∈ Rc×h×w, where c, h, and w denote the number of channels, height, and width, respectively, the operation yields an output feature map Y ∈ Rh′×w′×n, where n is the number of output channels and h′ and w′ are the height and width, respectively, of the resulting feature map. Accordingly, the FLOPs required for the standard Conv and GhostConv operations are defined as follows:
In this context, both the standard Conv and GhostConv operations employ convolutional filters with kernels size of k × k, whereas the cheap operation uses depthwise convolution filters with sizes of d × d. In this study, d was assumed to be equal to or approximately equal to k. Consequently, the theoretical compression ratio achieved by GhostConv is expressed as follows:
3.3.2. MLCA Module
To strengthen the feature representation capability of the network, a lightweight MLCA module is incorporated [35].
As shown in Figure 4, the MLCA module compresses the spatial dimensions of the input feature map through sequential local average pooling (LAP) and global average pooling (GAP) operations. The pooled features are then processed by a 1D convolution to highlight the features of the important channels. Moreover, another local enhancement path is processed by a 1D convolution after the feature reconstruction stage, which maintains the local structural information. After being fused through elementwise addition, the features derived from both paths undergo upsampling of non-local average pooling (UNAP) to achieve complementarity between local and global contexts. Finally, a multiplication operation applies the calculated weights to the original features, emphasizing critical information and achieving a more refined channel attention effect.
Figure 4.
MLCA module structure.
3.3.3. DGFPN Module
To fuse the feature representations acquired from different levels or sources and address the image feature loss caused by grayscale feature degradations, thereby enhancing the discriminative power of the network and its ability to represent and understand input data, the RepGFPN module, which efficiently aggregates features from various levels, and an ultralightweight and efficient dynamic upsampler called DySample are introduced in this study. Together, they constitute an enhanced feature fusion network, which is referred to as the DGFPN.
The structure of the RepGFPN module, which employs five CSPStage modules for integrating feature information obtained from different levels, is shown in Figure 5. This configuration effectively captures the multiscale features contained within coal and gangue images while enhancing the reusability of the features and the representation capability of the model. Furthermore, the recursive operations employed in the CSPStage modules strengthen the feature transmission and refinement processes, and the reparameterization mechanism contributes to reducing the inference time required by the model.
Figure 5.
RepGFPN module structure.
Moreover, the DySample module contained within the neck network mitigates the increases in parameter count and computational overhead caused by the RepGFPN module. As shown in Figure 6, the input feature map X is processed by a sampling point generator to produce a set of sampling coordinates. Here, the parameter s represents the spatial upsampling factor, while 2g corresponds to the 2-D coordinate components (x, y) of the original feature map X. Utilizing a predefined sampling function, the network enhances the spatial resolution of the input feature map X on the basis of the spatial position information acquired from the coordinate set, ultimately generating a high-resolution feature map X′.
Figure 6.
DySample module structure.
3.3.4. LSCD Module
The standard YOLOv8n architecture features three independent detection heads, which results in a significant number of parameters and high computational demand. To further reduce the weight of the model, An LSCD head is proposed. LSCD enables the sharing of a single set of detection kernel weights across feature maps with different scales. Specifically, it applies a 1 × 1 convolution for dimensionality reduction and channel alignment purposes across all features and then fuses the features through a 3 × 3 convolution operation. This composite optimization strategy achieves a balance between operational efficiency and recognition accuracy.
To mitigate the robustness reduction caused by the use of small batches, in the LSCD module structure shown in Figure 7, the modules are configured as Conv_GN modules, which combine group normalization (GN) with standard convolution operations. In addition, each Conv_Box module is followed by a scale module for size scaling, enabling the network to effectively accommodate targets with varying scales.
Figure 7.
LSCD module structure.
3.3.5. PIoUv2 Loss Function
Under the harsh conditions of coal and gangue detection scenarios, these materials exhibit complex texture variations and color attributes. The complete intersection over union (CIoU) loss function employed by the standard YOLOv8n model suffers from the issue of anchor box expansion, which is detrimental in detection scenarios where coal and gangue objects frequently overlap. Therefore, the powerful IoUv2 (PIoUv2) loss function [36] is introduced to address these limitations. By incorporating a penalty factor P that adaptively adjusts to target dimensions, the PIoU loss function effectively resolves the issue of anchor box expansion. It is defined as follows:
where dw1, dw2, dh1 and dh2 represent the absolute distances between the corresponding edges of the prediction box and the ground-truth box, and wgt and hgt denote the width and height of the ground-truth box, respectively, as shown in Figure 8. By utilizing the dimensions of the ground-truth box as the denominator of the penalty term, the model prevents anchor box expansion, ensuring that the penalty value P is solely dependent on the size of the ground-truth box.
Figure 8.
Distance penalty factor.
The PIoU loss function is defined as follows:
The PIoUv2 loss serves as an extension of the PIoU loss, incorporating a nonmonotonic attention layer that is controlled by a single hyperparameter. The function is defined as follows:
In Equation (12), q represents the correction coefficient for anchor boxes with varying quality levels. When q = 1, the penalty factor becomes zero, signifying complete alignment between the anchor box and the ground-truth box. u(λq) is an attention function, where λ serves as a key hyperparameter. By introducing the attention function, PIoUv2 prioritizes medium-quality anchor boxes during the regression process, thereby mitigating the negative impact of low-quality anchors on gradient updates. The use of λ as a single control hyperparameter significantly simplifies the parameter tuning process.
4. Experiment and Results
4.1. Experimental Setup and Evaluation Metrics
4.1.1. Experimental Configuration
In the study, all model training work was conducted under identical environmental conditions to ensure a fair comparison. The hardware configuration consisted of an NVIDIA RTX 3060 GPU (12 GB). The software environment was built on the Linux operating system using Python 3.8 and the PyTorch 1.13.1 deep learning framework. To ensure determinism and strict reproducibility of the training pipeline, a fixed random seed (seed = 16) was initialized across all computational libraries. During the experimental pipeline, the augmented dataset was consistently loaded with this fixed seed to prevent data shuffling variations. The specific training hyperparameters are detailed in Table 1.
Table 1.
Training parameters.
4.1.2. Evaluation Metrics
The performance of the proposed model was comprehensively evaluated using precision (P), recall (R), mean average precision (mAP), number of parameters (Params), number of giga floating-point operations (GFLOPs), and the model size as evaluation indicators. These indicators provide a holistic assessment of both the detection accuracy and the computational efficiency of the network.
Precision refers to the proportion of correctly predicted coal and gangue samples among all samples predicted as being coal and gangue. Recall refers to the proportion of correctly predicted coal and gangue samples among all actual coal and gangue samples. Their calculation formulas are shown in Equations (13) and (14), respectively:
In these equations, TP represents true positives, referring to the samples that the model predicts as positive and are actually positive, indicating that coal is correctly identified. FP represents false positives, which refer to the samples that should be negative but are predicted as positive by the model, specifically gangue samples that are incorrectly predicted as coal. FN represents false negatives, which refer to the samples that should be positive but are predicted as negative by the model, meaning coal samples that are incorrectly predicted as gangue.
The average precision (AP) is defined as the area enclosed by the PR curve and the coordinate axes under different confidence thresholds, and it reflects the accuracy of the model in this category. The average AP value across all categories is called the mAP, which can be used to comprehensively measure the performance of the tested model in terms of predicting the target and location categories. The calculation formulas for these metrics are shown in Equations (15) and (16):
In these equations, n represents the total number of categories contained in the dataset, and APi is the AP value achieved for each category. In this study, two types of IoU thresholds, mAP@50 and mAP@50:95, were selected to calculate the mAP. IoU was used to measure the degree of overlap between the predicted box and the ground-truth box. An IoU value closer to 1 indicates higher detection precision for the corresponding model.
The number of parameters is a key metric for evaluating the complexity of a model. It represents the total number of trainable parameters within the model and is generally measured by the total number of weight parameters included in the network.
4.2. Factors Affecting the Coal and Gangue Identification Process
4.2.1. Comparison Among the Average Grayscale Values Observed for Coal and Gangue
To quantitatively analyze the image features of coal and gangue, 80 images each of coal and gangue were extracted from the dataset, and partial samples are shown in Figure 9. After the mean grayscale value for each image was calculated, the data were plotted, as shown in Figure 10. As shown in the figure, the average grayscale values of gangue were mostly higher than those of coal. This was due to the significant differences in the visual characteristics of the two materials. Coal is primarily composed of carbonaceous organic matter, which appears deep black, has a high absorption rate and possesses a low visible light reflectance rate. Gangue is the associated inorganic waste generated during mining, which is lighter in color and has a higher visible light reflectance rate. Therefore, when a coal and gangue detection network is designed, the observed grayscale value differences can provide certain detection support. However, many images had average grayscale values of approximately 40, indicating that performing detection based solely on grayscale values is inaccurate.
Figure 9.
Sample images of coal and gangue. (a) Coal samples; (b) Gangue samples.
Figure 10.
Distribution of coal and gangue average grayscale value.
In actual working environments, factors such as illumination, the moisture state, and pulverized coal affect the identification of coal and gangue. To improve the achieved identification accuracy, it is necessary to analyze the grayscale characteristics of coal and gangue under different environmental factors.
4.2.2. Effects of Illumination Factors on the Grayscale Values of Coal and Gangue
To clarify the effects of illumination factors on the grayscale values of coal and gangue, samples were collected under three different illumination conditions: bright lighting, normal lighting, and dark lighting. On the basis of Equation (2), the variations exhibited by the grayscale histograms of coal and gangue under different illumination environments were calculated, as shown in Figure 11. The grayscale histograms reveal that as illumination intensity increased, the grayscale values of both coal and gangue exhibited overall rightward shifts, indicating that the overall brightness level improved. Additionally, the numbers of peaks contained in the histograms decreased, and the distribution ranges tended to become more dispersed. These findings indicate that illumination changes significantly affect the grayscale characteristics of coal and gangue.
Figure 11.
Grayscale histograms of coal and gangue under different illumination conditions.
Because coal has a high light absorption rate, local reflections are likely to occur when the illumination level increases, leading to uneven changes in the grayscale values of coal. In contrast, owing to the light color and strong light reflection characteristics of gangue, the overall change in its grayscale values is relatively stable. Therefore, the effects of illumination factors must be fully considered to improve the adaptability and robustness of coal and gangue detection methods. In addition, it is worth noting that the histograms of the other five samples overlap greatly, except for the coal image under dark lighting conditions, whose peak significantly differs. This grayscale overlap is a major reason why the conventional object detection methods find it difficult to distinguish between coal and gangue based solely on grayscale values.
4.2.3. Effect of the Moisture State on the Grayscale Values of Coal and Gangue
The coal and gangue samples were moistened and compared with dry samples, and the normalized grayscale histograms are plotted in Figure 12. After the coal and gangue were moistened, their grayscale values exhibited an overall leftward shift. This change was primarily caused by the impact of the moisture state on the surface optical properties of the materials. The presence of moisture enhances the light absorption rate, thereby reducing the reflection of visible light, which leads to a decrease in the overall grayscale values. Additionally, the formation of a water film may cause a specular reflection effect, altering the reflection characteristics of certain areas and thus affecting the grayscale distribution. The infiltration of water changes the color depth of coal and gangue; in the wet state, the overall color of the coal and gangue becomes darker, which further alters the grayscale histograms.
Figure 12.
Grayscale histograms of coal and gangue under different moisture state.
Environmental factors, such as illumination and moisture state conditions, affect the grayscale characteristics of coal and gangue to varying degrees, resulting in features such as grayscale overlapping and texture degradation. During coal and gangue detection, the utilized detection methods should be adjusted on the basis of field environmental fluctuations and the summarized patterns of the feature variations exhibited under interference. This approach not only improves the adaptability of the model in complex environments but also reduces the impact of single-factor interference on the resulting identification performance.
4.3. Training and Experimental Results of YOLOv8n-Coalgangue
4.3.1. Comparison Experiments Involving Different Backbone Networks
To verify the effectiveness of HGNetV2-G as the backbone network, the detection results produced by YOLOv8n variants based on different mainstream backbone networks on the dataset were compared, as shown in Table 2. CSPDarkNet [37] is the original backbone network of YOLOv8n. In the coal and gangue detection task, its mAP@50 and mAP@50:95 reached 95.3% and 70.3%, respectively, with a parameter count of approximately 3.1 M and a computational workload of 8.1 GFLOPs. By replacing the different backbone networks, the current mainstream backbone networks, such as MobilenetV4 [38], EfficientVit [39] and RepVit [40], were proven ineffective in these coal and gangue detection tasks and neither improved the resulting accuracy nor reduced the number of required parameters and calculations. Among the lightweight backbones, HGNetV2, GhostNet and StarNet [41] reduced the number of parameters and calculations. Among them, GhostNet had the smallest number of parameters and calculations—only 1.72 MB and 5.0 GFLOPs, respectively—but the accuracy impact of this method was more serious. Compared with those of YOLOv8n, the mAP@50 and mAP@50:95 of this method decreased by 0.3% and 1.3%, respectively. HGNetV2 performed slightly better, with decreases in only 0.2% and 0.6%, respectively.
Table 2.
Comparison of training results for different backbone networks.
Among these backbone networks, HGNetV2-G achieved the best performance. Its parameter count and computational workload were 2.31 M and 6.8 GFLOPs, respectively, and it delivered superior accuracy in the coal and gangue object detection task, with mAP@50 and mAP@50:95 values reaching 95.5% and 70.3%, respectively. The experiments demonstrated that using HGNetV2-G as the backbone network satisfies both the lightweight design and guaranteed accuracy requirements imposed in coal and gangue object detection tasks.
4.3.2. Comparison Experiments Involving Different Attention Mechanisms
To compare the effects of different attention mechanisms on the performance of the detection model, various attention mechanisms were integrated into the neck network for comparative experiments. Four types of attention mechanisms—EMA [42], the CBAM [43], ELA [44], and CAA [45]—were selected to construct improved models for performance comparison purposes. The detailed results are shown in Table 3.
Table 3.
Comparison of training results for different attention mechanisms.
The training results showed that when the five attention mechanisms were added to the neck, the performance of MLCA was the best, with its mAP@50 and mAP@50:95 values reaching 95.9% and 70.7%, respectively. Compared with those of the model using HGNetV2-G as the backbone, both metrics improved by 0.4%, while the parameter count remained almost unchanged. The other four attention mechanisms were significantly outperformed by the MLCA module; only the EMA module showed slightly better performance in terms of the mAP@50:95 metric, but it fell short of the MLCA module with respect to all the other indicators.
The experimental results demonstrate that integrating the MLCA mechanism into the neck can significantly enhance the feature representation ability of the utilized network, thereby improving the accuracy and effectiveness of the constructed model. In complex industrial and mining environments, especially when the surface of gangue is heavily covered by pulverized coal, its overall light reflectance rate changes significantly, leading to severe global grayscale aliasing with actual coal. Traditional attention mechanisms rely mainly on global average pooling and are prone to failures because of the degradation of global low-frequency grayscale features, thereby resulting in false detections. In contrast, the performance advantages of the MLCA module are significant. The local average pooling branch integrated within it effectively preserves local spatial structural information and forces the network, under globally dimmed conditions, to focus on the local high-frequency features that are not completely obscured, such as geometric discontinuities at gangue edges and underlying rough surface textures. This mechanism enables the constructed model to avoid relying solely on global grayscale cues that are prone to failure and instead break through visual camouflage by amplifying local texture and structure differences, thereby achieving the high-precision recognition of coal and gangue even under extreme interference scenarios such as coal dust accumulation.
4.3.3. Comparison Experiments Involving Different Loss Functions
To evaluate the performance of the PIoUv2 loss function in terms of model training and overall performance, it was compared with the CIoU of the original YOLOv8n and the current mainstream loss functions, such as the DIoU [46], EIoU [47], WIoU [48], SIoU [49], and PIoU losses [50]. All comparative experiments concerning the loss functions were conducted on the basis of the structurally improved YOLOv8 network model. The training results obtained using the different loss functions are shown in Table 4.
Table 4.
Training results using different loss functions.
When the improved model structure used the original CIoU, its mAP@50 and mAP@50:95 reached 96.6% and 73.3%, respectively. With the exception of the EIoU and WIoU loss functions, all loss functions outperformed the CIoU loss function of the original YOLOv8n model. Among them, the PIoU yielded the highest accuracy (94.9%), representing an increase of 0.7% over that of the original loss function. Furthermore, PIoUv2 delivered the best performance in terms of precision, with its mAP@50 and mAP@50:95 values reaching 97.2% and 73.9%, respectively, which were 0.6% higher than those of the original loss function, and the accuracy was improved by 0.4%, reaching 94.6%. Therefore, the comparative experiments involving different loss functions demonstrate that for the coal and gangue object detection task, the introduction of PIoUv2 yielded the best results and achieved the highest precision in the improved YOLOv8n model.
4.4. Ablation Study
To understand the impact of each improved component contained in the network on the resulting model performance and to evaluate the extent of these influences, an ablation study was conducted on the proposed improvements in this paper. The results of the ablation study are presented in Table 5. Among them, the first line indicates that no improvement modules have been added, and the check mark indicates the content of the improved modules added. The findings indicate that the five improved components based on YOLOv8n were effective in the coal and gangue object detection task, with the detection accuracy increasing by 0.2%, 0.3%, 0.3%, 0.2%, and 0.7%, respectively, compared with the original model.
Table 5.
Results of the ablation experiments.
Further experiments revealed that the combination of these improved components produced no negative effects; instead, it synergistically enhanced the performance of the model. First, replacing the backbone with HGNetV2-G reduced the parameter count to only 2.31 M while maintaining high feature extraction capability, increasing the resulting accuracy by 0.2%. Second, the introduction of the MLCA mechanism enhanced the robustness of the model in complex environments, increasing its accuracy to 95.9%, which was 0.6% higher than that of the original model. Next, changing the feature fusion network to the DGFPN improved the ability of the model to capture detailed features, further increasing its accuracy to 1.1% above that of the original model, which verifies the effectiveness of the dynamic fusion strategy designed for addressing grayscale feature degradations. Furthermore, the detection head was changed to LSCD to reduce the parameter quantity, which is a complex problem caused by the DGFPN for feature fusion. Thus, the parameter quantity of the model was reduced to 2.1 M, which was 32% lower than that of the original model. The model size was only 4.6 MiB, and the accuracy of the model was improved to 96.6%. Finally, replacing the loss function with PIoUv2 accelerated the convergence of the model and further increased its accuracy to 97.2%, representing a total improvement of 1.9% over the original model.
The data changes in Table 5 clearly indicate that the multiple improved components proposed in this paper could jointly achieve a lightweight model without reducing its accuracy. First, the HGNetV2-G backbone reduced the number of parameters from 3.1 M to 2.31 M, establishing a lightweight foundation. MLCA and the DGFPN jointly captured image features, where MLCA removed noise without increasing the parameter burden, whereas the DGFPN significantly improved the accuracy to 96.6% but also increased the number of parameters to 2.74 M, reflecting the cost of acquiring richer features. To solve this problem, LSCD formed a deep collaboration with the DGFPN, maintaining 96.6% accuracy without losing the feature representation ability of the model, while reducing the number of parameters to 2.1 M. With PIoUv2 optimizing the regression process for overlapping targets, this data-driven strategy enabled the model to achieve 97.2% accuracy with a size of only 4.6 MiB, thus satisfying the requirements of edge deployment in industrial environments.
4.5. Comparative Experiments
To further verify the performance of the YOLOv8n-CoalGangue network in coal and gangue detection tasks, experiments were conducted under the same experimental environment to compare this network with YOLOv10n [51] and YOLO11n [52]. The experimental results are shown in Table 6. To eliminate the influence of randomness during training and verify the statistical reliability of the outcomes, all results were obtained from 10 independent runs conducted under identical hyperparameter settings with different random seeds.
Table 6.
Comparison experiments with other network models.
As shown in the table, the YOLOv8n-CoalGangue network outperformed YOLOv8n, YOLOv10n, and YOLOv11 in terms of overall performance. With respect to accuracy, the developed model achieved an mAP@50 of 97.2 ± 0.15%, with a 95% confidence interval of [97.11, 97.28], indicating stable convergence performance. The mAP@50:95 of the YOLOv8n-CoalGangue network reached 74.5%. These results were superior to those of the other algorithms, with mAP@50:95 values that were 4.2%, 6.9%, and 4.7% higher, respectively, than those of the other three networks.
In terms of lightweight design, the proposed algorithm had a parameter count of 2.1 M and a computational workload of 5.7 GFLOPs, which was only 53.3% of the computational workload of YOLOv9-tiny and also significantly lower than those of the other two network models. When the data type was FP16, the model size of YOLOv8n-CoalGangue was only 4.6 MiB, which represents reductions of 1.5 MiB, 1.2 MiB, and 0.9 MiB, respectively, relative to the other networks, demonstrating superior performance. When these data points were synthesized, it was evident that the proposed improved YOLOv8n algorithm maintained higher precision while remaining as lightweight as possible.
4.6. Visualization of Results
To intuitively observe the feature extraction ability of the network model, heatmap visualization analysis was conducted on the coal and gangue test images using the EigenCAM algorithm. Furthermore, to contrast the improvement effects provided by the YOLOv8n-CoalGangue network, the heatmap results of the original YOLOv8n network were also included for comparison purposes. The experimental results are shown in Figure 13.
Figure 13.
Comparison of heat maps before and after model improvement.
According to Figure 13, the YOLOv8n-CoalGangue model clearly outperformed the original model in terms of detection performance. First, with respect to spatial localization accuracy, the high-response regions of the original YOLOv8n exhibited noticeable boundary diffusion effects, with attention extending beyond the actual contours of the coal and gangue. However, the improved model strictly constrained its attention within the true boundaries of the targets. Second, in terms of complete coverage of multiple targets, in scenarios with scattered gangue fragments, the improved model maintained strong and uniform attention across all target regions, effectively overcoming the limitation regarding the weak attention paid by the original model to some small fragments. Most importantly, in terms of suppressing background false positives, the original model produced many meaningless yellow–green activation responses in the background regions, which were highly likely to trigger false detections; in contrast, the background regions of the improved model exhibited an extremely clean deep-blue low-response state. This visual evidence intuitively and convincingly demonstrates the fact that YOLOv8n-CoalGangue has significant advantages with respect to effectively separating complex background interference, fully capturing target features, and accurately delineating contours.
4.7. Robustness Experiments Conducted in Different Scenarios
To evaluate the robustness of YOLOv8n-CoalGangue across diverse conditions, images derived from six complex scenarios were selected for testing. These conditions included dark environment, piles of pulverized coal, dense targets, occluded targets, motion blur, and high noise levels. These scenarios represent the various challenges that are typically encountered during the coal and gangue detection process. The detection results produced by the model are shown in Figure 14, where yellow boxes indicate missed detections and green boxes indicate false alarms.
Figure 14.
Robustness experiment results of YOLOv8n-CoalGangue under different scenarios. (a) Original image; (b) The effect of YOLOv8n detection; (c) The effect of YOLOv8n-CoalGangue detection. The yellow and green dashed ellipses highlight the false detection areas and missed detection areas, respectively.
The experimental results demonstrate that in dark lighting environments containing eight pieces of coal, the YOLOv8n model missed two coal targets. In contrast, the YOLOv8n-CoalGangue model successfully detected all coal pieces with high precision. In the scenario involving piles of coal dust, where the image contained two pieces of coal and one piece of gangue, the YOLOv8n model produced a false alarm by misidentifying a coal dust pile as a piece of coal. However, the YOLOv8n-CoalGangue model avoided this error entirely. In the dense-target scenario with six pieces of coal and two pieces of gangue, the YOLOv8n model missed one coal target in the crowded area on the right side of the image, whereas the YOLOv8n-CoalGangue model achieved complete detection of all targets. In the occluded-target scenario where eight pieces of coal and five pieces of gangue were stacked together, the YOLOv8n model produced two false alarms, whereas the YOLOv8n-CoalGangue model successfully detected all targets. In the motion blur scenario containing five pieces of coal, the YOLOv8n model missed one coal target because two adjacent pieces of coal were blurred during movement. However, the YOLOv8n-CoalGangue model completely detected all targets under these conditions. In the high-noise scenario featuring ten pieces of coal and five pieces of gangue, the YOLOv8n model missed one piece of gangue and misidentified another piece as coal at the bottom of the image. In contrast, the YOLOv8n-CoalGangue model accurately and completely detected all targets.
In terms of the results obtained across these six scenarios, the YOLOv8n model consistently suffered from missed detections and false alarms. In contrast, compared with the baseline model, the YOLOv8n-CoalGangue model avoided these issues entirely while maintaining superior detection precision.
To further evaluate the generalizability of the proposed model beyond a single dataset, the public DsCGF dataset [53] was introduced for external validation purposes. Samples from the Wangjialing Coal Mine were selected for testing. Without any fine-tuning on the external dataset, the model still achieved an mAP@50 of 93.6%. Although a slight performance decrease was observed because of domain shifts between different mining environments, these results demonstrate that the proposed model driven by grayscale characterization did not overfit a single dataset and could capture transferable grayscale features of coal and gangue.
5. Research Significance and Discussion
The core research significance of this study is the introduction of a solution driven by grayscale characterization for conducting machine vision-based detection under complex and harsh conditions. From a theoretical perspective, most existing deep learning-based coal and gangue detection methods follow a data-driven paradigm and treat the constructed model as a black box. They often rely on stacking parameters to fit features. In contrast, this study quantitatively analyzes the dynamic shifts and degradations exhibited by grayscale distributions under uneven illumination conditions and severe coal dust occlusion. The results show that global grayscale aliasing is the key factor that leads to model failures.
From engineering and economic perspectives, efficiently detecting coal and gangue is essential for conducting green mining and attaining clean coal utilization. The proposed YOLOv8n-CoalGangue framework achieves 97.2% accuracy with only 2.1 M parameters through integration of HGNetV2-G, LSCD, and other components. This lightweight design addresses the limitations of edge devices in terms of their computational complexity and power consumption levels. It enables real-time and low-cost dry sorting to be performed on belt conveyors. Moreover, the model provides a useful reference for other industrial vision tasks involving limited resources and harsh environments.
However, although the proposed model is strongly robust to moisture and illumination state changes, it still relies on residual surface texture information. This is a clear limitation. Under extremely high-moisture conditions, where the target is fully covered by coal slurry, texture information may be completely lost. In such cases, coal and gangue become visually indistinguishable, and the visual detection process based only on RGB features reaches its limit. The MLCA mechanism cannot recover information that no longer exists.
To address this limitation, future work should move beyond single-modality approaches. One direction is to introduce multimodal sensing systems to compensate for degraded visual features. Another direction is to incorporate unsupervised domain adaptation into the lightweight framework. This would allow the model to self-calibrate and continuously adapt during long-term deployment scenarios.
6. Conclusions and Future Work
This study addresses the challenge of automatically detecting coal and gangue in complex mining environments. By analyzing the degradation mechanism of grayscale features under environmental interference, a lightweight detection framework driven by grayscale characterization, i.e., YOLOv8n-CoalGangue, is proposed. A quantitative analysis shows that illumination variations and moisture conditions cause significant grayscale degradation, and severe coal dust occlusion leads to global grayscale aliasing. These findings provide important grayscale characterization priors for model design purposes. The experimental results and theoretical analysis demonstrate that the model, with its HGNetV2-G backbone and LSCD module, effectively reduces the number of redundant computations caused by homogeneous background features. Under a lightweight configuration with a 32% reduction in the number of parameters (2.1 M) and 5.7 GFLOPs, the model achieves 97.2% detection accuracy and a 126 FPS inference speed. This satisfies the strict requirements of edge devices in coal preparation plants. For challenging conditions such as coal dust occlusion cases, the effectiveness of MLCA and DGFPN is verified. These modules suppress low-frequency background noise and compensate for multiscale details, which helps overcome grayscale aliasing issues. In addition, the PIoUv2 loss function improves the bounding-box regression process in dense and overlapping scenarios through its angle and distance penalty designs.
Although the model is strongly robust under various extreme conditions, its performance still depends on the availability of surface texture information. When targets are fully covered by high-moisture slurry, visual features may be completely lost, and the visual detection process reaches its limit. Therefore, future work will involve exploring multimodal sensing techniques and integrating unsupervised domain adaptation to further improve the intelligent coal sorting procedure.
When deployed in practical applications, the proposed YOLOv8n-CoalGangue network not only significantly reduces the hardware investment costs incurred by edge devices but also alleviates the pressure imposed on the traditional wet coal washing processes. Saving industrial water resources and reducing wastewater discharge volumes provide considerable economic and environmental benefits.
Author Contributions
Methodology, G.Z.; software, W.X.; validation, G.Z., W.X. and Q.W.; formal analysis, Q.Z.; investigation, G.Z.; resources, Q.Z.; data curation, G.Z. and W.X.; writing—original draft preparation, G.Z. and W.X.; writing—review and editing, W.X.; visualization, Q.W.; supervision, Z.M.; project administration, Z.M.; funding acquisition, Z.M. All authors have read and agreed to the published version of the manuscript.
Funding
This research was funded by the National Natural Science Foundation of China (Grant No. 52474175 and U23A20599); Shandong Provincial Higher Education Institutions Youth Innovation Team Development Plan and the Open Fund Project of the Joint Collaborative Innovation Center for Mine Intelligent Technology and Equipment (Grant No. XTZX202403).
Data Availability Statement
The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author.
Acknowledgments
The authors extend their appreciation to the reviewers for their valuable feedback.
Conflicts of Interest
The author Guangyu Zhou was employed by the company Yili No. 1 Coal Mine, Shandong Energy Group Xinjiang Energy & Chemical Co. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
References
- Sui, Y.; Zhang, L.; Sun, Z.; Yi, W.; Wang, M. Research on Coal and Gangue Recognition Based on the Improved YOLOv7-Tiny Target Detection Algorithm. Sensors 2024, 24, 456. [Google Scholar] [CrossRef] [Scilit]
- Li, J.; Wang, J. Comprehensive utilization and environmental risks of coal gangue: A review. J. Clean. Prod. 2019, 239, 117946. [Google Scholar] [CrossRef] [Scilit]
- Long, M.; Schafrik, S.; Kolapo, P.; Agioutantis, Z.; Sottile, J. Equipment and Operations Automation in Mining: A Review. Machines 2024, 12, 713. [Google Scholar] [CrossRef] [Scilit]
- Xue, B.; Zhang, Y.; Li, J.; Wang, Y. A review of coal gangue identification research−application to China’s top coal release process. Environ. Sci. Pollut. Res. 2023, 30, 14091–14103. [Google Scholar] [CrossRef] [Scilit]
- Luo, X.; He, K.; Zhang, Y.; He, P.; Zhang, Y. A review of intelligent ore sorting technology and equipment development. Int. J. Miner. Met. Mater. 2022, 29, 1647–1655. [Google Scholar] [CrossRef] [Scilit]
- Ni, C.; Xu, G.; Chang, J.; Liu, B. Dense medium cyclone separation of fine coal: A discussion on the separation lower limit. Minerals 2023, 13, 1115. [Google Scholar] [CrossRef] [Scilit]
- Li, J.; Yang, D.; Du, C. Evaluation of an underground separation device of coal and gangue. Int. J. Coal Prep. Util. 2023, 33, 188–193. [Google Scholar] [CrossRef] [Scilit]
- Shang, D.; Wang, Y.; Yang, Z.; Wang, J.; Liu, Y. Study on comprehensive calibration and image sieving for coal-gangue separation parallel robot. Appl. Sci. 2023, 10, 7059. [Google Scholar] [CrossRef] [Scilit]
- Liu, P.; Tian, H.; Cao, X.; Qiao, X.; Gong, L.; Duan, X.; Qiu, Y.; Su, Y. Pick–and–Place Trajectory Planning and Robust Adaptive Fuzzy Tracking Control for Cable–Based Gangue–Sorting Robots with Model Uncertainties and External Disturbances. Machines 2022, 10, 714. [Google Scholar] [CrossRef] [Scilit]
- An, X.; Wang, Z.; Wei, D.; Gu, J.; Li, F.; Zhang, C.; Xia, G. YOLO-DFBL: An Improved YOLOv11n-Based Method for Pressure-Relief Borehole Detection in Coal Mine Roadways. Machines 2026, 14, 150. [Google Scholar] [CrossRef] [Scilit]
- Zhou, W.; Wang, H.; Wang, L.; Feng, A.; Li, L.; Zhu, J. Research on separation mechanism of coal gangue photoelectric sorting recognition. Physicochem. Probl. Miner. Process. 2023, 60, 192417. [Google Scholar] [CrossRef] [Scilit]
- Yin, J.; Zhu, J.; Zhu, H.; Pan, G.; Zhu, W.; Zeng, Q.; Shi, Q. Intelligent photoelectric identification of coal and gangue—A review. Measurement 2024, 233, 114723. [Google Scholar] [CrossRef] [Scilit]
- Luo, Q.; Wang, S.; Guo, Y.; He, L.; Cheng, G.; Li, X. Foreground segmentation and location of coal and gangue under complex similar background. Meas. Sci. Technol. 2023, 34, 035406. [Google Scholar] [CrossRef] [Scilit]
- Li, N.; Wu, S.; Yu, Z.; Gong, X. Feature extraction with multi-fractal spectrum for coal and gangue recognition based on texture energy field. Nat. Resour. Res. 2023, 32, 2179–2195. [Google Scholar] [CrossRef] [Scilit]
- Cheng, G.; Wei, Y.; Chen, J.; Pan, Z. A sorting method for coal and gangue based on surface grayness and glossiness. Gospod. Surowcami Miner. 2023, 39, 173–198. [Google Scholar] [CrossRef] [Scilit]
- Luo, Q.; Wang, S.; Guo, Y.; He, L.; Li, X. Adaptive image enhancement and particle size identification method based on coal and gangue. Meas. Sci. Technol. 2023, 34, 105403. [Google Scholar] [CrossRef] [Scilit]
- Gao, R.; Sun, Z.; Li, W.; Pei, L.; Hu, Y.; Xiao, L. Automatic coal and gangue segmentation using U-Net based fully convolutional networks. Energies 2023, 13, 829. [Google Scholar] [CrossRef] [Scilit]
- Ye, T.; Chen, H.; Ren, H.; Zheng, Z.; Zhao, Z. LPT-Net: A Line-Pad Transformer Network for efficiency coal gangue segmentation with linear multi-head self-attention mechanism. Measurement 2023, 226, 114043. [Google Scholar] [CrossRef] [Scilit]
- Li, N.; Zhang, A.; Han, H.; Duan, Y. RCPU-Net: A multi-scale multi-object segmentation model for coal and gangue under uneven lighting based on improved U-Net. Digit. Signal Process. 2026, 168, 105484. [Google Scholar] [CrossRef] [Scilit]
- Qin, Z.; Jing, J.; Li, L.; Yuan, Y.; Li, Y.; Li, B. Research on Image Segmentation and Defogging Technique of Coal Gangue Under the Influence of Dust Gradient. Appl. Sci. 2025, 15, 1947. [Google Scholar] [CrossRef] [Scilit]
- Si, L.; Chen, M.; Li, J.; Liu, Y.; Wang, Z.; Li, X.; Shao, H.; Wei, D. Visual image-based coal-gangue release status recognition with Detail-WaveNet and lightweight YOLOv8n model. Meas. Sci. Technol. 2025, 36, 106121. [Google Scholar] [CrossRef] [Scilit]
- Wang, L.; Wang, X.; Li, B. Data-driven model SSD-BSP for multi-target coal-gangue detection. Measurement 2025, 219, 113244. [Google Scholar] [CrossRef] [Scilit]
- Gao, J.; Li, B.; Wang, X.; Zhang, J.; Wang, L. A coal and gangue detection method for low light and dusty environments. Meas. Sci. Technol. 2025, 35, 035402. [Google Scholar] [CrossRef] [Scilit]
- Yang, Y.; Li, D.; Guo, Y.; Wang, S.; Zhao, D.; Chen, W.; Zhang, H. Research on coal gangue recognition method based on XBS-YOLOv5s. Meas. Sci. Technol. 2025, 35, 015404. [Google Scholar] [CrossRef] [Scilit]
- Gao, L.; Yu, P.; Dong, H.; Wang, W. Multi-Scale Fusion Lightweight Target Detection Method for Coal and Gangue Based on EMBS-YOLOv8s. Sensors 2025, 25, 1734. [Google Scholar] [CrossRef] [Scilit]
- Cao, X.; Liu, H.; Liu, Y.; Li, J.; Xu, K. Coal and gangue detection networks with compact and high-performance design. Sensors 2025, 24, 7318. [Google Scholar] [CrossRef] [Scilit] [PubMed]
- He, L.; Wang, S.; Guo, Y.; Cheng, G.; Hu, K.; Zhao, Y.; Wang, X. Multi-scale coal and gangue dual-energy X-ray image concave point detection and segmentation algorithm. Measurement 2022, 196, 111041. [Google Scholar] [CrossRef] [Scilit]
- Zhao, Y.; He, L. Research on small target segmentation method in dual-energy X-ray images of coal and gangue based on APSO-C_Otsu. Int. J. Coal Prep. Util. 2024, 44, 1447–1458. [Google Scholar] [CrossRef] [Scilit]
- Yan, P.; Wen, Z.; Wu, Z.; Li, G.; Zhao, Y.; Wang, J.; Wang, W. Intelligent detection of coal gangue in mining Operations using multispectral imaging and enhanced RT-DETR algorithm for efficient sorting. Microchem. J. 2024, 207, 111789. [Google Scholar] [CrossRef] [Scilit]
- Yan, P.; Wang, W.; Li, G.; Zhao, Y.; Wang, J.; Wen, Z. A lightweight coal gangue detection method based on multispectral imaging and enhanced YOLOv8n. Microchem. J. 2024, 199, 110142. [Google Scholar] [CrossRef] [Scilit]
- Li, H.; Wang, Q.; Ling, L.; Lv, Z.; Liu, Y.; Jiao, M. Research on recognition of coal and gangue based on laser speckle images. Sensors 2024, 23, 9113. [Google Scholar] [CrossRef] [Scilit] [PubMed]
- Elias, M.; Castiglione, P.; Elias, G. Influence of interface roughness on surface and bulk scattering. J. Opt. Soc. Am. A 2024, 27, 1265–1273. [Google Scholar] [CrossRef] [Scilit]
- Zhao, Y.; Lv, W.; Xu, S.; Wei, J.; Wang, G.; Dang, Q.; Liu, Y.; Chen, J. DETRs Beat YOLOs on Real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 17–21 June 2024; pp. 16965–16974. [Google Scholar] [CrossRef] [Scilit]
- Han, K.; Wang, Y.; Tian, Q.; Guo, J.; Xu, C.; Xu, C. GhostNet: More Features from Cheap Operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 14–19 June 2020; pp. 1577–1586. [Google Scholar] [CrossRef] [Scilit]
- Wan, D.; Lu, R.; Shen, S.; Xu, T.; Lang, X.; Ren, Z. Mixed local channel attention for object detection. Eng. Appl. Artif. Intell. 2023, 123, 106442. [Google Scholar] [CrossRef] [Scilit]
- Liu, C.; Wang, K.; Li, Q.; Zhao, F.; Zhao, K.; Ma, H. Powerful-IoU: More straightforward and faster bounding box regression loss with a nonmonotonic focusing mechanism. Neural Netw. 2024, 170, 276–284. [Google Scholar] [CrossRef] [Scilit]
- Terven, J.; Córdova-Esparza, D.-M.; Romero-González, J.-A. A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS. Mach. Learn. Knowl. Extr. 2023, 5, 1680–1716. [Google Scholar] [CrossRef] [Scilit]
- Qin, D.; Leichnner, C.; Delakis, M.; Fornoni, M.; Luo, S.; Yang, F.; Wang, W.; Banbury, C.; Ye, C.; Akin, B.; et al. MobileNetV4: Universal models for the mobile ecosystem. In Proceedings of the 18th European Conference on Computer Vision (ECCV), Milan, Italy, 29 September–4 October 2024; pp. 78–96. [Google Scholar] [CrossRef] [Scilit]
- Cai, H.; Li, J.; Hu, M.; Gan, C.; Han, S. EfficientViT: Lightweight Multi-Scale Attention for High-Resolution Dense Prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Paris, France, 1–6 October 2023; pp. 17256–17267. [Google Scholar]
- Wang, A.; Chen, H.; Lin, Z.; Han, J.; Di, G. RepViT: Revisiting Mobile CNN from ViT Perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 16–22 June 2024; pp. 15909–15920. [Google Scholar] [CrossRef] [Scilit]
- Ma, X.; Dai, X.; Bai, Y.; Wang, Y.; Fu, Y. Rewrite the Stars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 16–22 June 2024; pp. 5694–5703. [Google Scholar] [CrossRef] [Scilit]
- Ouyang, D.; He, S.; Zhang, G.; Luo, M.; Guo, H.; Zhan, J.; Huang, Z. Efficient Multi-scale Attention module with cross-spatial learning. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Rhodes, Greece, 4–9 June 2023; Available online: https://api.semanticscholar.org/CorpusID:258535361 (accessed on 6 November 2025).
- Woo, S.; Park, J.; Lee, J.; Kweon, I. CBAM: Convolutional Block Attention Module. In Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 8–14 September 2018; Volume 11211, pp. 3–19. [Google Scholar] [CrossRef] [Scilit]
- Xu, W.; Wan, Y.; Zhao, W. ELA: Efficient location attention for deep convolution neural networks. J. Real-Time Image Process. 2025, 22, 140. [Google Scholar] [CrossRef] [Scilit]
- Cai, X.; Lai, Q.; Wang, Y.; Wang, W.; Sun, Z.; Yao, Y. Poly kernel inception network for remote sensing detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 16–22 June 2024; pp. 27706–27716. [Google Scholar] [CrossRef] [Scilit]
- Zheng, Z.; Wang, P.; Liu, W.; Li, J.; Ye, R.; Ren, D. Distance-IoU Loss: Faster and better learning for bounding box regression. In Proceedings of the 34th AAAI Conference on Artificial Intelligence/32nd Innovative Applications of Artificial Intelligence Conference/10th AAAI Symposium on Educational Advances in Artificial Intelligence, New York, NY, USA, 7–12 February 2020; Volume 34, pp. 12993–13000. [Google Scholar] [CrossRef] [Scilit]
- Zhang, Y.; Ren, W.; Zhang, Z.; Jia, Z.; Wang, L.; Tan, T. Focal and efficient IOU loss for accurate bounding box regression. Neurocomputing 2022, 506, 146–157. [Google Scholar] [CrossRef] [Scilit]
- Chen, M.; Zhang, M.; Peng, J.; Huang, J.; Li, H. A Multi-Category Defect Detection Model for Rail Fastener Based on Optimized YOLOv8n. Machines 2025, 13, 511. [Google Scholar] [CrossRef] [Scilit]
- Wang, H.; Kou, Z.; Wang, Y. GES-YOLO: A Light-Weight and Efficient Method for Conveyor Belt Deviation Detection in Mining Environments. Machines 2025, 13, 126. [Google Scholar] [CrossRef] [Scilit]
- Chen, Z.; Chen, K.; Lin, W.; See, J.; Yu, H.; Ke, Y.; Yang, C. PIoU Loss: Towards accurate oriented object detection in complex environments. In Proceedings of the European Conference on Computer Vision (ECCV), Glasgow, UK, 23–28 August 2020; Volume 12350, pp. 195–211. [Google Scholar] [CrossRef] [Scilit]
- Liao, L.; Song, C.; Wu, S.; Fu, J. A Novel YOLOv10-Based Algorithm for Accurate Steel Surface Defect Detection. Sensors 2025, 25, 769. [Google Scholar] [CrossRef] [Scilit]
- Ali, M.L.; Zhang, Z. The YOLO Framework: A Comprehensive Review of Evolution, Applications, and Benchmarks in Object Detection. Computers 2024, 13, 336. [Google Scholar] [CrossRef] [Scilit]
- Lv, Z.; Fan, Y.; Sha, T.; Cui, Y.; Wu, Y.; Lv, H.; Sun, M.; Tu, Y.; Xu, Z.; Wang, W. A large-scale open image dataset for deep learning-enabled intelligent sorting and analyzing of raw coal. Sci. Data 2025, 12, 403. [Google Scholar] [CrossRef] [Scilit] [PubMed]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license.













