A real-time detection method for sonar images based on target detection neural network

Abstract

The present invention discloses a method for real-time detection of sonar images based on a target detection neural network. The present invention comprises the following steps: step (1), constructing a sonar image dataset and dividing it into a training set and a test set; step (2), data preprocessing: performing noise reduction, data normalization, and full-resolution data augmentation processing on the constructed training set and test set; step (3), constructing an underwater sonar image real-time detection network; step (4), training the sonar image real-time detection network based on the training set; step (5), verifying the accuracy and real-time performance of the sonar image real-time detection network based on the test set. The present invention draws on the framework of networks such as the YOLO series, SSD, and RetinaNet, designs and uses a backbone network, and maximizes the amount of input information. The method increases the sample size by data expansion, and improves the method by methods such as noise reduction preprocessing and loss function design to meet the accuracy requirements of the detection network. Finally, a SonarDet system is designed that improves prediction accuracy while maintaining the speed advantage.

Description

Technical Field

[0001] The present invention belongs to the field of ocean image detection, and in particular relates to a sonar image real-time detection method based on a target detection neural network. Background Art

[0002] As the economic and defense value of the ocean grows, the effective utilization of marine resources and the protection of maritime rights and interests are becoming increasingly urgent. Improving sonar target detection technology is imperative. The United States has placed particular emphasis on automatic sonar target recognition technology, investing significant resources and manpower, and achieving remarkable results. However, the research on sonar target recognition technology is complex and long-term, and requires constant monitoring of sonar equipment development. This has led to a long period of slow development. To date, sonar target recognition remains both a difficult and hot topic in research. Due to the late start of sonar target detection and recognition research in China, the country lags behind developed countries such as the United States. Therefore, the development of technology that can accurately detect and identify underwater targets is of paramount importance.

[0003] Currently, the best way to detect and identify underwater targets is to use sonar data to generate images of the seafloor. However, sonar data isn't as intuitive as optical images, and manual observation is inefficient. Therefore, the sonar target detection process requires a matching target recognition algorithm and system. Automatic sonar target recognition not only improves target detection efficiency but also provides timely feedback on detection results, meeting the requirements of modern military applications requiring real-time monitoring and strikes at sea. With the rapid increase in the demand for improving my country's naval combat capabilities in recent years, strengthening research on sonar target recognition technology has become increasingly important.

[0004] With the widespread adoption of deep learning models, many researchers have begun applying them to underwater target recognition and detection, aiming to address the shortcomings of traditional statistical machine learning methods. In 2013, Kamal S proposed a deep learning method based on a deep belief network (DBN) to classify multiple underwater targets. Results showed a classification accuracy of 90.23% for a 40-category classification problem. Eric L. Ferguson et al. proposed applying convolutional neural networks to underwater vessel noise recognition in shallow water environments. Experimental results showed that this network can identify targets over 180 meters away, a significant improvement over traditional statistical learning methods. Valdenegro-Toro M et al. applied convolutional neural networks to underwater sonar target detection, using underwater acoustic images as training samples. Experimental results showed that target classification accuracy exceeded 90%.

[0005] Models based on deep neural networks have powerful nonlinear feature learning capabilities and can implement end-to-end processing using raw data as input. Therefore, this project will use deep neural networks as a technical approach. Targeted improvements to deep neural networks to adapt them for the specific application scenario of sonar image target detection have great theoretical significance.

[0006] This project is a collaboration with the National Key Laboratory of Underwater Acoustic Technology. In the late 1990s, the laboratory conducted research on vector acoustic sensors and their applications, combining imported technology with independent innovation. This pioneered new areas of underwater acoustic technology, developing a variety of sensors with independent intellectual property rights, obtaining four national patents, and successfully applying them to various underwater acoustic monitoring systems. This has enabled my country to become one of the few developed countries to master this technology.

[0007] The Key Laboratory of Underwater Acoustics pioneered the research and development of acoustic sensors in China, becoming a key base for independent innovation in new transducer technologies and leading a wave of research in this field among domestic underwater acoustics peers. In recent years, with support from the "211 Project" and supporting infrastructure development, the laboratory has completed a number of high-level scientific research projects, significantly enhancing the laboratory's infrastructure and competitiveness. The laboratory has achieved remarkable results in the research of key underwater acoustic technologies, many of which are at the international advanced level.

[0008] Our research team has completed a deep neural network model based on Faster RCNN coupled with difficult sample burst training and used it for target detection tasks in sonar images.

[0009] However, this model still requires generating candidate regions of interest (ROIs) in sonar images during the data processing phase. Therefore, we sought to use a purely end-to-end deep network for learning to reduce the workload at this stage. Building on Ross Girshick's residual learning principles and drawing on networks such as the YOLO family, RetinaNet, and SSD, we designed our SonarDet network using a one-stage architecture and made several improvements to better suit sonar target detection. Preliminary research results show no significant difference in measurement results compared to the previous phase. Summary of the Invention

[0010] The purpose of the present invention is to provide a real-time detection method of sonar images based on target detection neural network in response to the deficiencies of the existing technology.

[0011] This paper proposes a one-stage SonarDet network based on Ross Girshick's residual learning idea and draws on the framework of networks such as the YOLO series, SSD, and RetinaNet. The present invention provides a real-time sonar image detection method based on a target detection neural network, comprising the following steps:

[0012] Step (1), constructing a sonar image dataset and dividing it into a training set and a test set;

[0013] Step (2), data preprocessing: noise reduction, data normalization, and full-resolution data augmentation are performed on the constructed training set and test set;

[0014] Step (3), constructing a real-time detection network for underwater sonar images, including a backbone network, a multi-scale fusion module, a kmeans algorithm module, a target detection module and a Loss function design module; the backbone network selects an improved darknet-53 network, and the images in the pre-processed training set are processed by the backbone network to obtain two high-channel-dimension low-resolution feature maps of different scales; the multi-scale fusion module fuses the obtained feature maps of the two scales to obtain a fused feature map; the kmeans clustering module is used to cluster the annotation boxes in the training set data to obtain 6 mother prior boxes; multiple prior boxes are randomly generated according to the proportion and position of the mother prior boxes; the target detection module predicts the fused feature maps of the two scales based on the prior boxes obtained by the kmeans clustering module, and identifies the sonar target position in the fused feature map by predicting the four offsets of the predicted box coordinates relative to the prior box;

[0015] Step (4), training the sonar image real-time detection network based on the training set;

[0016] Step (5) verifies the accuracy and real-time performance of the sonar image real-time detection network based on the test set.

[0017] Furthermore, step (1) is specifically implemented as follows:

[0018] The sonar image dataset comes from the dataset provided by China Shipbuilding Industry Corporation 715. The dataset is divided into a training set and a test set at an 8:2 ratio, and the training set data is annotated. The training set contains 4002 images and the test set contains 1000 images. The target detection object categories include columnar and linear shapes.

[0019] Furthermore, the specific structure of the backbone network in step (3) is as follows:

[0020] The backbone network uses the first 52 layers of darknet-53, that is, removing the fully connected layers and retaining the residual path and LeakyReLU activation function;

[0021] The size of the preprocessed image is 1024*1024*1, and it is input into the backbone network: first, it enters a convolution kernel with 32 filters, and the output image size is 1024*1024*32, and then it is input into 5 groups of repeated residual units at a time. Each residual unit of these 5 groups of residual units consists of a separate convolution layer and a group of repeated convolution layers. The repeated convolution layers are repeated 1 times, 2 times, 8 times, 8 times, and 4 times respectively; in each repeated convolution layer, a 1×1 convolution operation is performed first, and then a 3×3 convolution operation is performed. The number of filters is halved and then restored, for a total of 52 layers; finally, the feature maps of the last two layers of the backbone network are selected and output separately to obtain two feature maps with high channel dimensions and low resolution of different scales.

[0022] Furthermore, the kmeans algorithm module in step (3) is specifically implemented as follows:

[0023] The fused feature map is first divided into a nine-square grid, with two scales in each grid and three mother prior boxes at each scale. A total of two thousand prior boxes are randomly generated based on the proportions and positions of these 54 mother prior boxes. The six mother prior boxes in each grid are the same. The kmeans algorithm module counts the size and aspect ratio of the annotated boxes of the sonar images in the training set and clusters them into six categories.

[0024] Furthermore, the Loss function design module in step (3) is specifically implemented as follows:

[0025] In the category and cross entropy section, the Focal Loss function is used to solve the problem of severe imbalance in the ratio of positive and negative samples in one-stage target detection. An acceleration factor α is added to the Focal Loss function. Therefore, the specific formula of the final Loss function is as follows:

[0026]

[0027] Among them, λ coord is a hyperparameter used to balance the influence of coordinate loss, λ cls is a hyperparameter used to balance the influence of classification loss, K represents the number of network output layer grids, M represents the number of prior boxes, represents the degree of responsibility between the jth prior frame and the detection target in the i-th grid, β is the hyperparameter used to amplify the coordinate loss of the small frame, x i and y i Respectively represent the horizontal and vertical coordinates of the center point of the i-th annotation box, and Represents the horizontal and vertical coordinates of the center point of the i-th prediction box, ω i and h iRepresent the width and height of the i-th annotation box, and They represent the width and height of the i-th prediction box respectively, and FocalLoss represents the FocalLoss function, which is used to calculate its classification loss.

[0028] Furthermore, the acceleration factor α is designed as follows:

[0029] First, the white point value of any area in the dataset image is defined as the average value of each element in the pixel array of that area. Given that the white point value of most target areas in sonar images is higher than that of the background, the acceleration factor α can bypass the network's calculation of additional loss to affect the calculation of the gradient, thereby accelerating the network's convergence. The specific formula of the acceleration factor α is as follows:

[0030]

[0031] x = average pixel value

[0032] In the formula of the acceleration factor α, a is a hyperparameter used to increase the penalty strength, that is, to balance the value distribution, and x represents the average pixel value. The input value of the acceleration factor α is the white point value x of the prediction box generated by the network. The output value range is (1, a], which is negatively correlated with the input value.

[0033] The beneficial effects of the present invention are as follows:

[0034] 1. This paper adopts a one-stage fully convolutional network framework to achieve real-time target detection in sonar images.

[0035] 2. The present invention designs the SonarDet network, designs and uses the backbone for this project, and improves the accuracy of the detection network through the application of multi-scale fusion technology and kmeans clustering algorithm and targeted design of loss function, so as to enable it to detect the target.

[0036] 3. The present invention makes targeted improvements to the loss function and designs an acceleration factor α, which accelerates the convergence of the network by affecting the gradient calculation of the loss function, improves the accuracy of the network, and realizes the real-time performance of the network.

[0037] 4. The present invention adopts a cosine annealing learning rate reduction method during SonarDet network training. By suddenly increasing the learning rate, it can "jump out" of the local minimum and find a path to the global minimum, thereby improving the accuracy of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is the SonarDet network structure diagram;

[0039] Figure 2 For sonar images;

[0040] Figure 3 This is a comparison chart before and after noise reduction processing;

[0041] Figure 4 This is the flow chart of the underwater sonar image real-time detection system;

[0042] Figure 5 This is the test result sonar AP result graph;

[0043] Figure 6 This is the result graph of the test result sonar F1;

[0044] Figure 7 This is the result graph of the test result sonar accuracy;

[0045] Figure 8 This is the result diagram of the test result Recall;

[0046] Figure 9 is the test result diagram;

[0047] Figure 10 This is the result graph of the test result Ground truth;

[0048] Figure 11 This is the result graph of the test result Log average miss rate;

[0049] Figure 12 Bit test result mAP value; DETAILED DESCRIPTION

[0050] The present invention will be further described below with reference to the accompanying drawings.

[0051] This paper proposes a real-time underwater sonar image detection system based on target detection neural network. Based on the idea of residual learning of Ross Girshick, and drawing on the framework of networks such as YOLO series, SSD and RetinaNet, a one-stage SonarDet network is designed. The flow chart of the system is shown in the figure. Figure 4 As shown in the figure, the specific design steps of the underwater sonar image real-time detection system are as follows:

[0052] Step (1): Based on the sonar image dataset, establish a sonar image dataset for training and testing

[0053] The dataset of the present invention comes from the dataset provided by CSIC 715. The dataset is divided into a training set and a test set according to the ratio of 8:2, and the training set data is annotated. The training set contains 4002 images and the test set contains 1000 images. The original images are as follows: Figure 2As shown in the figure, the target detection object categories include columnar and linear shapes.

[0054] Step (2), perform data preprocessing: noise reduction, data normalization, and full-resolution data augmentation;

[0055] The original sonar images are preprocessed. Because the original sonar images are blurry and noisy, with a small sample size and image channels that differ from traditional image channels, data annotation and target detection network learning may be difficult. Therefore, the images are subjected to denoising, normalization, and data augmentation. The specific preprocessing steps are as follows:

[0056] 2-1 Noise Reduction Processing

[0057] Gaussian blur is used to remove the noise background and reduce the impact of noise on feature extraction to obtain higher quality images. While maintaining the integrity of the original information as much as possible, useless information in the signal is removed to enhance the subsequent detection effect. The comparison before and after noise reduction is as follows Figure 3 shown.

[0058] 2-2 Normalization Processing

[0059] Normalization is used to standardize the 16-bit source input data to eliminate the dimension effect, making it easier for the network to extract effective features.

[0060] 2-3 Data augmentation processing

[0061] Sonar samples are more difficult to obtain than natural images, and the number of samples is relatively small. Therefore, we use some data augmentation methods to increase the sample size to more fully train the deep network.

[0062] Step (3), constructing the underwater sonar image real-time detection network SonarDet, including the backbone network, multi-scale fusion module, kmeans algorithm module, target detection module and Loss function design module;

[0063] The flow chart of underwater sonar image real-time detection system is as follows: Figure 4 As shown in the figure, the input sonar image is sent to the SonarDet network after noise reduction, normalization, and full-resolution image preprocessing, and the output result is finally obtained. The design steps of the SonarDet network are as follows:

[0064] 3-1. The backbone network of the real-time sonar image detection network uses the first 52 layers of darknet-53, that is, the fully connected layers are removed and the residual path and leaky ReLU activation function are retained.

[0065] like Figure 1As shown in the figure, after image preprocessing, the image size is 1024*1024*1, and it is input into the backbone network. It first enters a convolution kernel with 32 filters, and the output image size is 1024*1024*32. This image is then input into five sets of repeated residual units (residual blocks). Each of these five residual units consists of a single convolutional layer and a set of repeated convolutional layers, which are repeated once, twice, eight times, eight times, and four times, respectively. In each repeated convolutional layer, a 1×1 convolution operation is performed first, followed by a 3×3 convolution operation. The number of filters is halved and then restored, for a total of 52 layers. The preprocessed image is processed by the backbone network, resulting in two feature maps with high channel dimensions and low resolution at different scales.

[0066] 3-2. In order to reduce the amount of redundant calculations, the multi-scale fusion module fuses the feature maps of the two scales to obtain a fused feature map.

[0067] A specified number of layers are selected from the last few layers of the backbone network. The outputs of the higher-level layers are downsampled to the output size of the lower-level layers. Finally, the outputs are added together to fuse the features. Since the sonar dataset has only two scales of annotation boxes, only two layers are selected to reduce the amount of redundant computation. The present invention calculates the scales of the annotation boxes of the dataset and finally selects the feature maps of the last two layers of the backbone network for output. The multi-scale fusion module fuses the feature maps of the two scales to obtain a fused feature map.

[0068] 3-3. To enhance the detection accuracy of the network, the kmeans clustering module is used to cluster the labeled boxes in the training set data to obtain 6 parent prior boxes. Multiple prior boxes are randomly generated based on the proportion and position of the parent prior boxes.

[0069] Because the real-time sonar image detection network detects targets based on prior frames, the prior frames are set by first dividing the fused feature map into a nine-square grid, with each grid having two scales and three parent prior frames at each scale. Two thousand prior frames are randomly generated based on the proportions and positions of these 54 parent prior frames. The six parent prior frames in each grid are identical. The kmeans algorithm is used to calculate the size and aspect ratio of the annotated boxes in the sonar images in the training set, clustering them into six clusters. This reduces the number of anchor frames and specifically modifies their aspect ratios to better suit the characteristics of the targets detected in sonar images and enhance recognition accuracy.

[0070] 3-4. The target detection module is implemented based on the prior box obtained by the kmeans clustering module. It predicts the fused feature map at two scales and identifies the sonar target position in the fused feature map by calculating the four offsets of the predicted box coordinates relative to the prior box.

[0071] 3-5. Loss function design module; During the network training process, since the underwater sonar image real-time detection network belongs to a one-stage network, there are problems such as too many simple negative samples and too large a loss ratio, resulting in low accuracy. The present invention has made a targeted design for the Loss function. This network uses the Focal Loss function in the category and cross entropy parts to solve the problem of serious imbalance in the ratio of positive and negative samples in one-stage target detection. On the basis of the Focal Loss function, the Loss function is further optimized according to the characteristics of sonar images, and an acceleration factor α is added. The specific formula of the Loss function of the present invention is as follows:

[0072]

[0073] In the Loss function, λ coord is a hyperparameter used to balance the influence of coordinate loss, λ cls is a hyperparameter used to balance the influence of classification loss, K represents the number of network output layer grids, M represents the number of prior boxes, represents the degree of responsibility between the jth prior frame and the detection target in the i-th grid, β is the hyperparameter used to amplify the coordinate loss of the small frame, x i and y i Respectively represent the horizontal and vertical coordinates of the center point of the i-th annotation box, and Represents the horizontal and vertical coordinates of the center point of the i-th prediction box, ω i and h i Represent the width and height of the i-th annotation box, and They represent the width and height of the i-th prediction box respectively, and FocalLoss represents the FocalLoss function, which is used to calculate its classification loss.

[0074] The dataset image preprocessed in step 2 is divided into 3×3 grids, where It means that after the data set is divided into S×S grids, the responsibility of the j-th prior box in the i-th grid to the detected target. That is, when the intersection-over-union ratio (IOU) between the j-th prior box in the i-th grid and the true value of the target is the largest among all prior boxes, the shape and size of this prior box best match the current target, then this prior box is used to predict the target. The rest are 0.

[0075] Since some of the prior boxes in this experiment are very small, the coordinate loss of the detected target is very small, so the hyperparameter β is set to subtract the coordinate loss, that is, β-ω i ×h i, to amplify the coordinate loss of the small frame. In the present invention, after multiple experiments, the parameter β is finally set to 2.

[0076] The final coordinate loss is calculated by first calculating the center point offset of each prior box and size offset Then multiply it by the coordinate loss coefficient, then add the coordinate loss of each box, and finally reference the FocalLoss function to calculate its classification loss and multiply it by the classification loss coefficient λ cls Then, it is added with the coordinate loss to obtain .

[0077] Based on the loss function obtained above, the present invention also designs an acceleration factor α. First, the white point value of any region in the dataset image is defined as the average value per element in the pixel array of that region. Given that the white point values of most target regions in sonar images are higher than those of the background, the acceleration factor α bypasses the network to calculate additional loss, thereby affecting the gradient calculation and accelerating the network's convergence. The specific formula for the acceleration factor α is as follows:

[0078]

[0079] x = average pixel value

[0080] In the formula of the acceleration factor α, a is a hyperparameter used to increase the penalty intensity, that is, to balance the value distribution, and x represents the average pixel value.

[0081] The input value of the acceleration factor α function is the white point value x of the predicted box generated by the network; the output value range is (1, a] and is negatively correlated with the input value. By modifying the function curve of the acceleration factor α, its impact on the loss can be controlled. According to the chain rule of gradient calculation, the loss obtained by multiplying the acceleration factor α by the original loss can exponentially and controllably increase the network parameter gradient calculated by backpropagation, effectively encouraging the network to converge to areas with high white point values when the learning rate is fixed.

[0082] Using this acceleration factor α can reduce the weight of easily classified samples, allowing the model to focus more on difficult-to-classify samples during training. This improvement allows the network to maintain the speed of a one-stage network while bringing the accuracy closer to that of a two-stage network, thereby improving the accuracy of the network.

[0083] Step (4), training the SonarDet network using the preprocessed sonar dataset;

[0084] Based on the sonar dataset preprocessed in step (2), the SonarDet neural network architecture is trained. Since the gradient descent algorithm used to optimize the objective function will make the loss value closer and closer to the global minimum, the present invention uses a cosine annealing learning rate descent method to "jump out" of the local minimum and find a path to the global minimum by suddenly increasing the learning rate. This descent mode can be combined with the learning rate to produce a very effective calculation method to produce good results and improve network accuracy.

[0085] Step (5) verifies the accuracy and real-time performance of the model.

[0086] The SonarDet model trained in step (4) is tested with the data test set to verify the accuracy and real-time performance of the model. After the test, the columnar and linear sonar AP results are as follows Figure 5 As shown, F1 Figure 6 As shown, the accuracy is Figure 7 As shown, the Recall value is Figure 8 As shown, the test results are Figure 9 As shown, Ground truth is as Figure 10 As shown, the Logaverage miss rate is Figure 11 As shown in Figure 12 shown.

Claims

1. A real-time detection method for sonar images based on target detection neural network, characterized in that The steps include: Step (1), constructing a sonar image dataset and dividing it into a training set and a test set; Step (2), data preprocessing: noise reduction, data normalization, and full-resolution data augmentation are performed on the constructed training set and test set; Step (3), constructing a real-time detection network for underwater sonar images, including a backbone network, a multi-scale fusion module, a kmeans algorithm module, a target detection module, and a Loss function design module; the backbone network selects an improved darknet-53 network, and the images in the pre-processed training set are processed by the backbone network to obtain two high-channel-dimensional low-resolution feature maps of different scales; the multi-scale fusion module fuses the obtained feature maps of the two scales to obtain a fused feature map; the kmeans clustering module is used to cluster the labeled boxes in the training set data to obtain 6 mother prior boxes; Multiple prior frames are randomly generated based on the proportion and position of the parent prior frame. The target detection module predicts the fused feature map of the two scales based on the prior frame obtained by the kmeans clustering module. The sonar target position in the fused feature map is identified by the four offsets of the predicted frame coordinates relative to the prior frame. Step (4), training the sonar image real-time detection network based on the training set; Step (5), verifying the accuracy and real-time performance of the sonar image real-time detection network based on the test set; Step (1) is specifically implemented as follows: The sonar image dataset comes from the dataset provided by China Shipbuilding Industry Corporation 715. The dataset is divided into a training set and a test set with an 8:2 ratio, and the training set data is annotated. The training set contains 4002 images and the test set contains 1000 images. The target detection object categories include columnar and linear shapes. The specific structure of the backbone network in step (3) is as follows: The backbone network uses the first 52 layers of darknet-53, that is, removing the fully connected layers and retaining the residual path and leaky ReLU activation function; The size of the preprocessed image is 1024*1024*1, and it is input into the backbone network: first, it enters a convolution kernel with 32 filters, and the output image size is 1024*1024*32, and then it is input into 5 groups of repeated residual units residualblock at a time. Each of these 5 groups of residual units consists of a separate convolution layer and a group of repeated convolution layers. The repeated convolution layers are repeated 1 times, 2 times, 8 times, 8 times, and 4 times respectively; in each repeated convolution layer, a 1×1 convolution operation is performed first, and then a 3×3 convolution operation is performed. The number of filters is halved and then restored, for a total of 52 layers; finally, the feature maps of the last two layers of the backbone network are selected and output respectively, and two feature maps with high channel dimensions and low resolution of different scales are obtained; The kmeans algorithm module in step (3) is specifically implemented as follows: The fused feature map is first divided into a nine-square grid, with each grid having two scales and three mother prior boxes at each scale. A total of 2,000 prior boxes are randomly generated based on the proportions and positions of these 54 mother prior boxes. The six mother prior boxes in each grid are identical. The kmeans algorithm module calculates the size and aspect ratio of the annotated boxes in the sonar images in the training set and clusters them into six categories. The Loss function design module in step (3) is specifically implemented as follows: In the category and cross entropy part, the Focal Loss function is used to solve the problem of severe imbalance in the ratio of positive and negative samples in one-stage target detection. An acceleration factor α is added to the FocalLoss function. Therefore, the specific formula of the final Loss function is as follows: Among them, λ coord is a hyperparameter used to balance the influence of coordinate loss, λ cls is a hyperparameter used to balance the influence of classification loss, K represents the number of network output layer grids, M represents the number of prior boxes, represents the degree of responsibility between the jth prior frame and the detection target in the i-th grid, β is the hyperparameter used to amplify the coordinate loss of the small frame, x i and y i Respectively represent the horizontal and vertical coordinates of the center point of the i-th annotation box, and Represents the horizontal and vertical coordinates of the center point of the i-th prediction box, ω i and h i Represent the width and height of the i-th annotation box, and Represent the width and height of the i-th prediction box respectively, FocalLoss represents the FocalLoss function, which is used to calculate its classification loss degree; The acceleration factor α is designed as follows: First, the white point value of any region in the dataset image is defined as the average value per element in the pixel array of that region. Given that the white point values of most target regions in sonar images are higher than those of the background, the acceleration factor α can bypass the network's calculation of additional loss to affect gradient calculation, thereby accelerating network convergence. The specific formula of the acceleration factor α is as follows: In the formula of the acceleration factor α, a is a hyperparameter used to increase the penalty intensity, that is, to balance the value distribution, and x represents the average pixel value.

Images