Pavement crack disease optimization identification method based on hierarchical mechanism and data enhancement

Abstract

The application discloses a pavement crack disease optimization identification method based on a grading mechanism and data enhancement, introduces a crack severity grading mechanism, effectively improves the detection accuracy of slight cracks, divides the disease detection into different grades, and provides clearer technical basis for subsequent core verification and engineering decision-making; on the basis of grading, a diversified data enhancement strategy is further introduced, training data sets are enriched through data expansion modes such as flipping, cropping, brightness adjustment, saturation adjustment and sharpening processing, the robustness and detection capability of the model to complex crack diseases are significantly improved, and the identification efficiency and accuracy of internal crack diseases of pavement structure in three-dimensional ground penetrating radar data processing are effectively improved.

Description

Technical Field

[0001] This application relates to the field of pavement distress identification and analysis technology, and in particular to an optimized identification method for pavement crack distress based on a hierarchical mechanism and data augmentation. Background Technology

[0002] Road surface crack detection is an important task in road maintenance and management. Traditional crack detection methods mainly rely on technologies such as manual inspection, laser scanning, and ultrasonic testing. However, with the increase in road inspection workload, the shortcomings of traditional methods have gradually become apparent. For example, manual inspection is inefficient and greatly affected by weather, lighting, and human factors. While laser scanning and ultrasonic testing technologies can provide high-precision crack information, their equipment is expensive, operation is complex, and they cannot process large-scale road data.

[0003] In recent years, with the development of deep learning and computer vision technologies, automatic detection methods based on images and radar signals have become a mainstream solution. Using convolutional neural networks (CNNs) for target detection, especially when combined with GPR (Ground Penetrating Radar) images and high-resolution images, can automatically and accurately detect and locate road surface cracks. Deep learning methods, through learning from large amounts of labeled data, can extract the spatial features of road surface cracks, thereby achieving efficient automated detection.

[0004] Currently, several deep learning-based pavement crack detection technologies have been applied to various scenarios. For example, a pavement crack detection method based on convolutional neural networks (CNNs) utilizes GPR images to identify cracks on the road surface. This method automatically identifies cracks by training a CNN model, but its accuracy and robustness are relatively low, especially in complex road environments where it faces difficulties in detecting small or deep cracks. Another related technology is based on YOLO series target detection models, proposing a YOLOv8-based pavement crack detection method that combines image recognition and multi-scale feature fusion techniques, effectively improving crack recognition accuracy. However, the ordinary YOLOv8 model still has limited accuracy and robustness when handling cracks of different depths and sizes, especially in different weather conditions, complex road conditions, and crack detection at different scales, where its detection performance falls short of ideal levels. Therefore, a joint optimization identification method based on the RTDETR-L model will become one of the important directions for improving pavement crack detection technology, providing strong technical support for the safety and sustainable development of infrastructure.

[0005] Existing crack detection technologies still face the following problems when dealing with cracks of varying depths, sizes, and backgrounds:

[0006] Impact of low-confidence data: Traditional confidence threshold grading methods usually filter out low-confidence data, but this data may contain important crack information, leading to missed detection problems.

[0007] Crack data with insufficient features: For some low-confidence data, it is often difficult to extract effective information directly from the raw data because the features are not obvious enough (such as insufficient lighting, occlusion, small cracks, etc.), resulting in poor detection results.

[0008] Insufficient diversity of training data: Although data augmentation techniques have been widely used, low-confidence data augmentation is still relatively weak in practical applications and has failed to fully improve the model's adaptability to complex cracks. Summary of the Invention

[0009] To address the aforementioned technical problems, this application provides an optimized identification method for pavement crack defects based on a hierarchical mechanism and data augmentation, overcoming the problem of insufficient generalization ability in existing models.

[0010] This application provides a method for optimizing the identification of pavement crack defects based on a hierarchical mechanism and data augmentation, including the following steps:

[0011] Step S1: Obtain a dataset of internal road surface defects based on three-dimensional ground-penetrating radar;

[0012] Step S2: Annotate and preprocess the dataset of internal cracks and defects in the pavement structure;

[0013] Step S3: Train the crack disease data based on the RTDETR-L model network structure and select the grading threshold to identify the crack data and classify the degree of severity.

[0014] Step S4: Perform data augmentation on data content in the classified dataset where the number of samples is less than a preset threshold;

[0015] Step S5 involves implementing a joint optimization strategy by combining confidence threshold grading with data augmentation methods, and then conducting experimental evaluation.

[0016] Furthermore, in step S1, the crack features involved in the data of the internal crack disease dataset of the pavement structure include at least the crack shape, size and depth.

[0017] Further, in step S3, in the RTDETR-L model network structure:

[0018] Features are extracted using a backbone network, and detection capabilities are enhanced by depthwise separable convolution and multi-scale feature fusion. The results are then output as high-precision target detection results through a decoder.

[0019] Furthermore, the confidence threshold is graded in the RTDETR-L model network structure. A class confidence value and the corresponding bounding box regression value are output for each predicted box, and a confidence score is generated.

[0020] Furthermore, the step of generating a confidence score by outputting a category confidence value and a corresponding bounding box regression value for each predicted box specifically includes:

[0021] In object detection tasks, confidence includes class confidence and bounding box confidence. In the RTDETR-L model, confidence is calculated based on the class and coordinate regression results of each predicted box, including class confidence and bounding box confidence. Class confidence P(class) represents the model's confidence that the object belongs to a specific class, and bounding box confidence P(bbox) represents the accuracy of the model in locating the predicted bounding box.

[0022] The confidence level calculation formula is shown below: ,

[0023] The confidence level of the detection box is set based on empirical values. Cracks with a confidence level below the confidence level are considered insignificant cracks, while cracks with a confidence level above the confidence level are considered significant cracks.

[0024] Furthermore, the data enhancement in step S4 includes image enhancement based on geometric transformation and image enhancement based on color transformation.

[0025] The geometric transformation-based image enhancement includes either flipping or cropping.

[0026] The color-transformation-based image enhancement includes at least one of brightness adjustment, saturation adjustment, and sharpening.

[0027] Furthermore, in step S5, the evaluation metrics include accuracy, recall, mean accuracy, and mean mean accuracy.

[0028] Accuracy represents the proportion of areas correctly identified as cracks out of all areas predicted as cracks in the detection results; accuracy ,in, For a true example, that is, a crack area that is correctly identified. A false positive, meaning an area that is mistakenly identified as a crack;

[0029] Recall rate represents the proportion of regions correctly identified as cracks out of all real crack regions in the detection results; ,in, These are false negatives, meaning crack areas that were not correctly identified.

[0030] Average accuracy is used to measure the overall recognition ability of a model at different confidence thresholds; average accuracy ,in, Indicates recall rate The corresponding accuracy, The higher the value, the better the overall performance of the model under different confidence thresholds;

[0031] Mean accuracy is used to evaluate the overall performance of multi-class object detection tasks; mean accuracy ,in, The total number of categories, Represents the j-th category value;

[0032] The beneficial effects of this invention are as follows: This application introduces a grading mechanism for the severity of cracks, which effectively improves the detection accuracy of minor cracks. At the same time, it divides crack defects into different levels, providing a clearer technical basis for subsequent core sampling verification and engineering decisions. Based on the grading, it further introduces diversified data augmentation strategies, enriching the training dataset through various data expansion methods such as flipping, cropping, brightness adjustment, saturation adjustment, and sharpening, which significantly improves the robustness and detection capability of the model for complex crack defects, and effectively improves the identification efficiency and accuracy of internal crack defects in pavement structures in 3D ground-penetrating radar data processing. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart of the method of the present invention;

[0035] Figure 2 This is the network structure design diagram of the RTDETR-L model of the present invention;

[0036] Figure 3 This is a structural design diagram of the HGBlock detection model module of the present invention;

[0037] Figure 4 This is a structural design diagram of the RepC3 module of the detection model described in this invention;

[0038] Figure 5 This is a comparison image of cracks after data enhancement as described in this invention;

[0039] Figure 6The images show the detection results before and after optimization as described in this invention. Numbers a, b, c, and d represent four different cracks. The numbers in the red boxes, such as crack0.83 and CA0.83, represent the model's confidence score. The model's confidence score is a core indicator that quantifies the model's degree of certainty regarding the correctness of its recognition results. The confidence score ranges from 0 to 1; a higher value indicates that the model believes there is a greater probability of a target crack within the recognition frame, and the higher the reliability of the result. The improved model exhibits significantly higher and more stable confidence, demonstrating higher recognition accuracy and reliability. Detailed Implementation

[0040] To make the purpose, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0041] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0042] In the description of this application, it should be understood that the terms "upper", "lower", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0043] This application uses the RTDETR-L model as a crack identification scheme to study the confidence threshold grading process for identifying internal pavement cracks, such as... Figure 1 As shown, it includes the following steps:

[0044] Step S1: Obtain a dataset of internal cracks and defects in the road surface structure based on three-dimensional ground-penetrating radar;

[0045] As a specific implementation, when acquiring a dataset of crack defects within the pavement structure, based on the characteristics of crack data, a 3D ground-penetrating radar device is used to acquire pavement crack image data and form a dataset. This dataset includes crack features of different types and depths. Considering that the morphology, size, and depth of cracks directly affect the model's recognition results, especially for smaller or shallower cracks, the model's detection confidence is usually low. Therefore, when subsequently selecting the grading threshold, the diversity of cracks and the model's ability to recognize different types of cracks need to be taken into account.

[0046] Step S2 involves preprocessing the dataset of internal cracks in the pavement structure and simultaneously optimizing the hyperparameters of the model. Unlike the parameters updated through backpropagation during model training, hyperparameters are externally exposed configuration items that require manual adjustment by the user based on experience or through repeated experiments to improve model performance. Specific parameters are shown in Table 1.

[0047] Table 1 Model Parameters

[0048]

[0049] Step S3: Based on the RTDETR-L model network structure, the confidence threshold is graded by outputting a category confidence value and the corresponding bounding box regression value for each predicted box, and a confidence score is generated.

[0050] Based on the confidence level data, the crack data are divided into obvious crack group and inconspicuous crack group;

[0051] As a specific embodiment, this application constructs an RT-DETR model network structure, which is an efficient and model-based target detection model.

[0052] When using this model:

[0053] Features are extracted using backbone networks (HGStem, HGBlock);

[0054] Improve detection capabilities by utilizing depthwise separable convolution and multi-scale feature fusion (such as upsampling and stitching);

[0055] The decoder outputs high-precision target detection results.

[0056] The backbone network consists of multiple modules used to extract features from the input image. These modules include:

[0057] HGStem: Initial convolutional layer used to generate preliminary features.

[0058] HGBlock: Extracts high-level features from an image, possibly through stacked convolutional operations (see attached). Figure 3 (As shown).

[0059] RepC3: An improved module based on the CSP architecture, used for efficient feature extraction and feature fusion (see attached). Figure 4 (As shown).

[0060] DWConv: Depthwise Convolution, used for efficiently processing spatially and channel-separated features. A stack of multiple HGBlock modules forms a deep feature extractor that progressively refines and abstracts the features of the input image.

[0061] In the RTDETR-L model, confidence is calculated based on the class and coordinate regression results of each predicted bounding box. Specifically, the class confidence P(class) represents the model's confidence that the target belongs to a specific class; while the bounding box confidence P(bbox) represents the accuracy of the model's localization of the predicted bounding box.

[0062] The model generates the final confidence score by outputting a class confidence score and the corresponding bounding box regression score (i.e., the box's location coordinates and size prediction) for each predicted box and then combining the two.

[0063] In target detection tasks, confidence score is a crucial metric for measuring the accuracy of detection boxes. Particularly in pavement internal crack detection, the generation of confidence scores not only affects the accuracy of the model's detection results but also directly impacts the effectiveness of subsequent defect classification, screening, and model optimization strategies. To achieve more accurate crack detection, this study, based on the output of the RTDETR-L model, calculates the confidence score value for each detection box to generate probabilities associated with crack features, thereby guiding subsequent crack defect classification and optimization.

[0064] In object detection tasks, confidence is typically used to measure a model's confidence in the correctness of its bounding boxes. It mainly consists of two parts: class confidence and bounding box confidence. In the RTDETR-L model, confidence is calculated based on the class and coordinate regression results of each predicted bounding box. Specifically, the class confidence P(class) represents the model's confidence that the object belongs to a specific class; while the bounding box confidence P(bbox) represents the accuracy of the model's localization of the predicted bounding boxes. The model outputs a class confidence value and a corresponding bounding box regression value (i.e., the box's position coordinates and size prediction) for each predicted bounding box, and then combines these two to generate the final confidence score. The confidence score calculation formula is shown below:

[0065] In the formula, It is the category confidence level; It is the bounding box confidence score.

[0066] In this embodiment, based on the generated confidence level, this application introduces a classification method based on confidence level threshold to divide the crack data into two categories: obvious cracks (CA) and inconspicuous cracks (CB).

[0067] Based on the experimental results and practical application requirements, this study adopted the following two threshold settings:

[0068] (1) CA (Clear Crack): When the confidence value of the detection box is between 0.7 and 1.0, the crack is considered a clear crack. Such cracks have high detection accuracy and model recognition confidence, and are usually considered high-quality samples for model detection, and should be given priority. Clear cracks are usually large and clear cracks, and the model has a strong recognition ability on these samples.

[0069] (2) CB (Inconspicuous Crack): When the confidence value of the detection box is between 0.25 and 0.7, the crack is considered inconspicuous. These cracks have low confidence, and the model lacks confidence in their identification, potentially leading to significant errors or uncertainties. Therefore, these cracks require further optimization, such as data augmentation or model fine-tuning, to improve detection accuracy and robustness.

[0070] Step S4: Perform data augmentation on the data of the inconspicuous crack group to address the problem of relatively small sample size of inconspicuous cracks leading to an imbalance in crack data categories.

[0071] This application introduces geometric transformations for image enhancement, generating diverse training samples by transforming the spatial geometric properties of the image.

[0072] The geometric transformations here include flipping and clipping.

[0073] The flipping mechanism includes two forms: horizontal flipping and vertical flipping. Its main function is to simulate various possible arrangements of cracks by changing the image's orientation, thereby increasing the model's robustness. Specifically, horizontal flipping symmetrically flips the image along the vertical axis, while vertical flipping symmetrically flips the image along the horizontal axis.

[0074] Cropping generates new training samples by randomly selecting local regions from an image. This method not only enhances the model's ability to capture local features of cracks but also simulates local observation situations caused by device limitations or changes in viewpoint in real-world scenarios. The cropping operation typically selects a certain proportion of the image region as new input while maintaining the integrity of the annotation information. The cropped image contains local details of the cracks, contributing to higher accuracy in handling complex local scenes.

[0075] As another embodiment, this application may also introduce color transformation to enhance the robustness of the model to different lighting conditions by adjusting the color attributes of the image. Specifically, this includes:

[0076] Brightness adjustment simulates different lighting conditions by changing the overall brightness of the image, such as the crack imaging effect during the day, at night, or in cloudy weather. Adjusting the brightness helps the model adapt to various ambient lighting changes, thereby improving its detection performance in different scenarios.

[0077] Contrast enhancement, by adjusting the distribution range of pixel values ​​in an image, highlights its visual features, thereby strengthening the representation of cracks in the image. Contrast enhancement can effectively increase the difference between crack edges and the background, making it easier for the model to capture crack features.

[0078] Saturation adjustment enhances the visual characteristics of cracks by adjusting the color saturation of the image, making it easier for the model to distinguish cracks from the background. Saturation adjustment can enhance the visual information of cracks by increasing or decreasing color intensity, making the model more accurate in detecting cracks under different environmental conditions. A comparison image of cracks after data augmentation is attached. Figure 5 As shown.

[0079] Step S5 involves implementing a joint optimization strategy by combining confidence threshold grading with data augmentation methods, and then conducting experimental evaluation.

[0080] As a specific implementation, in order to correctly evaluate the accuracy of the dataset of internal structural cracks in the pavement, this step uses precision (P), recall (R), mean accuracy (AP), and mean mean accuracy (mAP) as evaluation metrics.

[0081] Accuracy represents the proportion of areas correctly identified as cracks out of all areas predicted as cracks in the detection results; accuracy ,in, For a true example, that is, a crack area that is correctly identified. A false positive, meaning an area that is mistakenly identified as a crack;

[0082] Recall rate represents the proportion of regions correctly identified as cracks out of all real crack regions in the detection results; ,in, These are false negatives, meaning crack areas that were not correctly identified.

[0083] Average accuracy is used to measure the overall recognition ability of a model at different confidence thresholds; average accuracy ,in, Indicates recall rate The corresponding accuracy, The higher the value, the better the overall performance of the model under different confidence thresholds;

[0084] Mean accuracy is used to evaluate the overall performance of multi-class object detection tasks; mean accuracy ,in, The total number of categories, Represents the j-th category value;

[0085] The following is a horizontal comparison of the key steps:

[0086] The original model (RTDETR-crack1673), as the unoptimized model, uses a single global confidence threshold to detect cracks, without considering the different severity levels of cracks (CA and CB classes) and without augmenting the dataset. This model performs well in overall detection accuracy and recall, but due to the lack of crack grading, it still has certain limitations in detecting low-confidence cracks (CB class), as shown in Table 2.

[0087] Table 2 Experimental Results of RTDETR-crack1673

[0088]

[0089] The crack severity classification model (RTDETR-CACB) introduces a crack severity classification mechanism, enabling the RTDETR-CACB model to categorize cracks into two types: obvious cracks (CA class) and inconspicuous cracks (CB class). This classification allows the model to more accurately identify obvious and inconspicuous cracks, thereby improving overall detection accuracy. Specific results are shown in Table 3.

[0090] Table 3 Experimental results of RTDETR-CACB

[0091]

[0092] As shown in the table above, the overall performance of the model after grading is relatively balanced, especially in the detection of obvious cracks (CA class), where it achieves high precision and recall values. However, its detection performance on inconspicuous cracks (CB class) is still insufficient. By introducing a confidence grading strategy, the model can more clearly distinguish the severity of cracks, but further optimization is still needed, especially in terms of the detection accuracy of low-confidence cracks.

[0093] The data-augmented model (RTDETR-CACB-enhance), which is the final model adopted in this application, is a further optimized version of the RTDETR-CACB-enhance model. Based on the original classification, it incorporates various data augmentation methods, such as image flipping and cropping, significantly improving the model's ability to detect low-confidence cracks (CB class). Through the addition of data augmentation strategies, the model exhibits stronger robustness in handling complex pavement cracks, as shown in Table 4.

[0094] Table 4. Experimental results of RTDETR-CACB-enhance

[0095]

[0096] The data-augmented RTDETR-CACB-enhance model achieved significant improvements across multiple performance metrics, particularly in the detection accuracy of inconspicuous cracks (CB class), improving by approximately 9.0%. Compared to the RTDETR-CACB model, RTDETR-CACB-enhance showed significant improvements in precision, recall, mAP@0.5, and mAP0.5:0.95, especially in the detection precision and recall of CB class cracks. This indicates that data augmentation not only enhances the overall robustness of the model but also enables it to better adapt to complex crack scenarios.

[0097] Therefore, as we can see from the above model data analysis, general pavement crack detection models use a single global confidence threshold for crack detection, without considering the classification of crack types, and without performing data augmentation on the sample data. Such models suffer from insufficient generalization ability, especially in handling complex pavement cracks with low accuracy. The optimized model of this invention, however, comprehensively incorporates a confidence threshold grading strategy and data augmentation methods, resulting in clearer differentiation of crack levels and significant improvements in key indicators such as detection accuracy and recall.

[0098] Specifically, the optimized RTDETR-CACB model effectively improves the detection accuracy of minor cracks by introducing a grading mechanism based on crack severity (CA and CB classes). It also categorizes crack defects into different levels, providing a clearer technical basis for subsequent core sampling verification and engineering decisions. Furthermore, the optimized RTDETR-CACB-enhance model further incorporates diverse data augmentation strategies on top of the grading, enriching the training dataset through various data expansion methods such as flipping and pruning. This significantly improves the model's robustness and detection capability for complex crack defects, effectively enhancing the efficiency and accuracy of identifying internal crack defects in pavement structures during 3D ground-penetrating radar data processing.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0100] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations (such as quantity, shape, position, etc.) can be made to the technical solutions of the present invention, and these equivalent transformations are all protected by the present invention.

Claims

1. A method for optimized identification of pavement crack defects based on a hierarchical mechanism and data augmentation, characterized in that, Includes the following steps: Step S1: Obtain a dataset of internal cracks and defects in the road surface structure based on three-dimensional ground-penetrating radar; Step S2: Annotate and preprocess the dataset of internal cracks and defects in the pavement structure; Step S3: Train the crack disease data based on the RTDETR-L model network structure and select the grading threshold to identify the crack data and classify the degree of severity. Step S4: Perform data augmentation on data content in the classified dataset where the number of samples is less than a preset threshold; Step S5: By combining confidence threshold grading with data augmentation methods, a joint optimization strategy is implemented and experimental evaluation is conducted. In step S3, within the RTDETR-L model network structure: The backbone network is used to extract features, and the detection capability is enhanced by depthwise separable convolution and multi-scale feature fusion. Then, the decoder outputs high-precision target detection results. In the RTDETR-L model network structure, the confidence threshold is graded by outputting a class confidence value and the corresponding bounding box regression value for each predicted box, and a confidence score is generated. The process of generating confidence by outputting a category confidence value and a corresponding bounding box regression value for each predicted box specifically includes: In object detection tasks, confidence includes class confidence and bounding box confidence. In the RTDETR-L model, confidence is calculated based on the class and coordinate regression results of each predicted box, including class confidence and bounding box confidence. The class confidence P(class) represents the model's confidence that the object belongs to a certain class, and the bounding box confidence P(bbox) represents the accuracy of the model in locating the predicted bounding box. The confidence level calculation formula is shown below: , Based on empirical values, a confidence level grading system is established for the detection box. Cracks with a confidence level below the grading value are considered insignificant cracks, while cracks with a confidence level above the grading value are considered significant cracks.

2. The method for optimized identification of pavement crack defects based on hierarchical mechanism and data augmentation according to claim 1, characterized in that: In step S1, the crack features involved in the data of the internal crack disease dataset of the pavement structure include at least the crack shape, size and depth.

3. The method for optimized identification of pavement crack defects based on hierarchical mechanism and data augmentation according to claim 1, characterized in that: The data enhancement in step S4 includes image enhancement based on geometric transformation and image enhancement based on color transformation; The image enhancement based on geometric transformation includes at least one of flipping and cropping; The color-transformation-based image enhancement includes at least one of brightness adjustment, saturation adjustment, and sharpening.

4. The method for optimized identification of pavement crack defects based on hierarchical mechanism and data augmentation according to claim 1, characterized in that: In step S5, the evaluation metrics include accuracy, recall, mean accuracy, and mean mean accuracy. Accuracy represents the proportion of areas correctly identified as cracks out of all areas predicted as cracks in the detection results; accuracy ,in, For a true example, that is, a crack area that is correctly identified. A false positive, meaning an area that is mistakenly identified as a crack; Recall rate represents the proportion of regions correctly identified as cracks out of all real crack regions in the detection results; ,in, These are false negatives, meaning crack areas that were not correctly identified. Average accuracy is used to measure the overall recognition ability of a model at different confidence thresholds; average accuracy ,in, Indicates recall rate The corresponding accuracy, The higher the value, the better the overall performance of the model under different confidence thresholds; Mean accuracy is used to evaluate the overall performance of multi-class object detection tasks; mean accuracy ,in, The total number of categories, Represents the j-th category value.

Images