Method and system for realizing sonar image segmentation network

Abstract

The invention provides a sonar image segmentation network implementation method and system, and the method comprises the steps: collecting a sonar image, and constructing a training data set; preprocessing the training data set, and extracting features of sonar images of the training data set; performing maximum pooling on the feature maps of different scales to obtain a spliced information map; the spliced information graph is subjected to convolution, and a fused multi-scale feature graph is generated; recording the high-resolution feature map from the encoder to obtain a first high-resolution feature map; transposing and convoluting the fused multi-scale feature map to obtain a second high-resolution feature map; fusing the first high-resolution feature map and the second high-resolution feature map to obtain a fusion result; carrying out the convolution of the fusion result, and obtaining a pixel-level prediction result; and distributing the pixel-level prediction result, and outputting an image. According to the method, through multi-scale feature fusion and high-resolution feature recovery, the segmentation precision of the sonar image is remarkably improved, target segmentation is more accurate, and information of different levels can be captured.

Description

Technical Field

[0001] The present invention belongs to the technical field of sonar image processing, and in particular, relates to a sonar image segmentation network implementation method and system, and more particularly, to a sonar image segmentation network based on multi-scale feature fusion and an implementation method thereof. Background Art

[0002] Sonar imaging technology is a technology that uses sound waves for detection and imaging, and is widely used in ocean exploration, underwater target identification and other fields.

[0003] Specifically, sonar images generate image information of target objects by analyzing the propagation and reflection of sound waves in the medium. Due to the low contrast and noise interference of sonar images, it faces great challenges in target recognition and image segmentation tasks.

[0004] There are two main technologies for segmenting sonar images. One is the traditional image segmentation method, such as threshold segmentation and edge detection, which mainly relies on the grayscale and edge information of the image. These methods do not work well when processing sonar images with low contrast and noise interference, and are easily affected by background noise, resulting in inaccurate segmentation results. The other is an image segmentation method based on deep learning, such as convolutional neural network, or CNN, which is widely used in image segmentation tasks and extracts image features through deep convolutional layers. However, when processing sonar images, CNN is often affected by the reduction in spatial resolution of feature maps, resulting in loss of detail information. Encoder-decoder networks such as U-Net pass high-resolution feature maps to the decoder through jump connections to restore image details. Although this method performs well in some tasks, when processing sonar images, the existing network structure still has shortcomings in multi-scale feature fusion.

[0005] The problems existing in the prior art can be summarized into the following three aspects.

[0006] First: Feature loss. In the traditional encoder-decoder network, the pooling operation in the encoder stage can extract high-level features, but it also leads to the loss of position information and details. This will significantly affect the segmentation accuracy of the target in the sonar image when processing sonar images, especially small targets and complex backgrounds.

[0007] Second: Insufficient fusion of multi-scale features. In the existing technology, although some methods try to restore image details through jump connections, there are still problems with the effective fusion of multi-scale features. The targets in sonar images may have different scales and shapes, which requires the network to be able to effectively fuse feature information of different scales.

[0008] Third: Noise interference. Sonar images are often interfered by noise, which makes the segmentation task more difficult. Existing methods often lack sufficient robustness and stability when dealing with noise effects, resulting in inaccurate segmentation results.

[0009] Patent document CN117953341A discloses a pathological image segmentation network model, method, device and medium. The network model provided by the scheme includes an encoder and a decoder; the encoder includes a multi-scale convolution attention module, which is used to extract feature information of different scales of the input image through a number of convolution kernels of different sizes, and pay attention to the association information between feature information of different scales from the spatial dimension, obtain global context information, and suppress interference information in the feature information through the global context information; the decoder includes a multi-scale feature fusion module, which is used to fuse feature information of different scales. The pathological image segmentation network model of the scheme can better capture feature information of different scales and reduce noise interference through learning global context association information, while being able to efficiently retain edge information and fine-grained information, and can perform pathological image segmentation work with a higher accuracy; however, it cannot solve the technical problems of feature loss and insufficient multi-scale feature fusion.

[0010] This problem needs to be solved urgently. Summary of the invention

[0011] In view of the defects in the prior art, the object of the present invention is to provide a sonar image segmentation network implementation method and system.

[0012] A sonar image segmentation network implementation method provided by the present invention includes:

[0013] Step S1: Collect sonar images and construct a training data set;

[0014] Step S2: preprocessing the training data set, and extracting features of the sonar images of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales;

[0015] Step S3: maximum pooling the feature maps of different scales and splicing them into a spliced ​​information map; convolving the spliced ​​information map to generate a fused multi-scale feature map; recording the high-resolution feature map from the encoder to obtain a first high-resolution feature map;

[0016] Step S4: transpose convolve the fused multi-scale feature map to obtain a second high-resolution feature map; fuse the first high-resolution feature map with the second high-resolution feature map to obtain a fusion result; convolve the fusion result to obtain a pixel-level prediction result;

[0017] Step S5: assign the pixel-level prediction result and output an image.

[0018] Preferably, in step S1, Gaussian noise is added to the training data set;

[0019] In the step S2, it includes:

[0020] Step S2.1: data augmentation of the training data set;

[0021] Step S2.2: extracting features of the sonar image of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales;

[0022] In step S2.1, the data enhancement methods include: rotation, translation and scaling;

[0023] In the step S2.2, the features of the sonar image of the training data set are extracted from three scales, and feature maps of three different scales are obtained accordingly; the three scales include: 3×3 convolution kernel, 5×5 convolution kernel and dilated convolution;

[0024] Preferably, in step S3, the three feature maps of different scales are pooled by 2×2 maximum pooling and spliced ​​into a spliced ​​information map; the spliced ​​information map is convolved by 1×1 to generate a fused multi-scale feature map;

[0025] The high-resolution feature map is an image directly extracted from the encoder;

[0026] In step S4, the fusion methods include: splicing, weighted fusion and adaptive fusion;

[0027] The weighted coefficients of the weighted fusion are obtained through network learning;

[0028] The adaptive fusion is to adjust the fusion weight according to the task objectives.

[0029] Preferably, in step S5, a softmax function is applied to generate a classification probability distribution, and then the pixel results are allocated to output an image;

[0030] The steps of allocating include:

[0031] Step A1: Collect the output of the model to obtain the category score corresponding to each pixel, i.e. logits;

[0032] Step A2: converting the category scores into probability distributions through a softmax function;

[0033] Step A3: Based on the probability distribution, the category index with the highest probability is obtained by taking the maximum value index, and then the pixels are assigned to the corresponding categories to form the final classification result and output the image.

[0034] A sonar image segmentation network implementation system provided by the present invention is characterized by comprising:

[0035] Module M1: Collect sonar images and build training data sets;

[0036] Module M2: preprocessing the training data set, and extracting features of the sonar images of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales;

[0037] Module M3: Maximum pooling of the feature maps of different scales, and splicing them into a spliced ​​information map; convolving the spliced ​​information map to generate a fused multi-scale feature map; recording the high-resolution feature map from the encoder to obtain a first high-resolution feature map;

[0038] Module M4: transpose convolution the fused multi-scale feature map to obtain a second high-resolution feature map; fuse the first high-resolution feature map with the second high-resolution feature map to obtain a fusion result; convolve the fusion result to obtain a pixel-level prediction result;

[0039] Module M5: distributes the pixel-level prediction results and outputs an image.

[0040] Preferably, in the module M1, Gaussian noise is added to the training data set;

[0041] The module M2 includes:

[0042] Module M2.1: Data augmentation of the training dataset;

[0043] Module M2.2: extracting features of the sonar image of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales;

[0044] In the module M2.1, the data enhancement methods include: rotation, translation and scaling;

[0045] In the module M2.2, the features of the sonar image of the training data set are extracted from three scales to obtain feature maps of three different scales; the three scales include: 3×3 convolution kernel, 5×5 convolution kernel and void convolution.

[0046] Preferably, in the module M3, the three feature maps of different scales are 2×2 max pooled and spliced ​​into a spliced ​​information map; the spliced ​​information map is 1×1 convolved to generate a fused multi-scale feature map;

[0047] The high-resolution feature map is an image directly extracted from the encoder;

[0048] In the module M4, the fusion methods include: splicing, weighted fusion and adaptive fusion;

[0049] The weighted coefficients of the weighted fusion are obtained through network learning;

[0050] The adaptive fusion is to adjust the fusion weight according to the task objectives.

[0051] Preferably, in the module M5, a softmax function is applied to generate a classification probability distribution, and then the pixel results are allocated and an image is output;

[0052] The steps of allocating include:

[0053] Module A1: Collect the output of the model and obtain the category score corresponding to each pixel, i.e. logits;

[0054] Module A2: convert the category scores into probability distribution through the softmax function;

[0055] Module A3: Based on the probability distribution, the category index with the highest probability is obtained by taking the maximum value index, and then the pixels are assigned to the corresponding categories to form the final classification result and output the image.

[0056] According to a computer-readable storage medium storing a computer program provided by the present invention, when the computer program is executed by a processor, the steps of the sonar image segmentation network implementation method are implemented.

[0057] An electronic device provided according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the sonar image segmentation network implementation method are implemented.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1. The present invention significantly improves the segmentation accuracy of sonar images through multi-scale feature fusion and high-resolution feature restoration, making target segmentation more accurate. In addition, the multi-scale feature fusion provided by the present invention can capture information at different levels and improve the segmentation effect.

[0060] 2. The network provided by the present invention can effectively restore the detail information in the sonar image, overcoming the shortcomings of the traditional method in detail restoration. Specifically, the present invention can enhance the detail restoration capability and restore more detail information through improved skip connection and upsampling technology.

[0061] 3. The present invention enhances the network's ability to handle noise interference, making the segmentation results more stable and reliable in complex environments. Through improved convolution technology and robustness training, it can effectively cope with noise interference and improve noise robustness.

[0062] 4. The improved network structure provided by the present invention can provide excellent segmentation effects in multiple sonar application scenarios, such as ocean detection and underwater target recognition, which makes the present invention have broad practical application prospects and improves the scope of application. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0064] Figure 1 A schematic diagram of an encoder-decoder sonar image segmentation network provided by the present invention;

[0065] Figure 2 Schematic diagram of the multi-scale convolution branch module provided by the present invention. DETAILED DESCRIPTION

[0066] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0067] In order to solve the problems of insufficient multi-scale feature fusion and noise interference, the present invention proposes an improved multi-scale feature fusion sonar image segmentation network, the main purpose of which is to solve the multiple technical challenges encountered in the prior art when processing sonar images. Specifically, the present invention aims to:

[0068] 1. Overcoming the problem of feature loss. In traditional encoder-decoder networks, due to the pooling operation in the encoder stage, the high-resolution detail information of the image is often lost, thus affecting the final segmentation result. According to the theory of CNN, the pooling operation reduces the spatial resolution of the feature map by downsampling, usually using maximum pooling or average pooling. Although this operation can reduce the computational complexity and increase the receptive field, it inevitably loses some spatial detail information, affecting the accurate recognition of the target.

[0069] 2. Solve the problem of insufficient fusion of multi-scale features. Although the encoder-decoder network in the prior art can restore high-resolution features through skip connections, it still has shortcomings in effectively fusing multi-scale features.

[0070] 3. Enhance the ability to handle noise interference. Sonar images are often interfered by noise, and existing methods show poor robustness when dealing with noise influence.

[0071] To achieve the above objectives, the improved multi-scale feature fusion sonar image segmentation network proposed in the present invention is implemented as follows:

[0072] 1. Provide multi-scale convolution branches. In the encoder stage, multiple parallel convolution branches are designed to extract local details, large-scale context information, and information about the expanded receptive field. Each branch uses convolution kernels of different scales, such as 3×3, 5×5, and dilated convolution.

[0073] 2. Feature map concatenation and fusion: The feature maps extracted by the multi-scale convolution branches are concatenated and then fused through 1×1 convolution.

[0074] 3. Provide skip connections. Save high-resolution feature maps of each stage in the encoder and pass these feature maps to the corresponding stage of the decoder through skip connections.

[0075] 4. Upsampling and fusion: In the decoder, these feature maps are upsampled using transposed convolution or interpolation methods.

[0076] The upsampled feature map is fused with the high-resolution feature map of the jump connection. The fusion strategies include: splicing fusion, weighted fusion and adaptive fusion.

[0077] 5. Improve the convolution method. Use noise suppression technology in the convolution layer and use the noise suppression module of the convolution filter to enhance the robustness to noise.

[0078] 6. Robust training: Use noisy data during training, adopt data augmentation techniques, and inject random noise to improve the robustness of the model.

[0079] The present invention is particularly a multi-scale feature fusion technology applied to sonar image segmentation tasks. Specifically, the present invention proposes an improved encoder-decoder network structure, which improves the segmentation accuracy and detail recovery ability of targets in sonar images by effectively fusing multi-scale features.

[0080] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

[0081] Figure 1 Schematic diagram of the encoder-decoder sonar image segmentation network based on multi-scale feature fusion.

[0082] A sonar image segmentation network implementation method provided by the present invention includes:

[0083] Step 1: Collect a representative forward-looking sonar image dataset A.

[0084] Step 2: Add Gaussian noise to the training data A in step 1 to enhance the robustness of the model.

[0085] Step 3: Use rotation, translation, scaling and other methods to perform data enhancement on the image obtained in step 2 to expand the training sample data set.

[0086] Step 4: Design multiple parallel convolution branches in the encoder, use convolution kernels of different scales, and each convolution branch extracts features of different scales. Figure 2 As shown in the figure, convolution branch 1 uses a 3×3 convolution kernel to extract local features, convolution branch 2 uses a 5×5 convolution kernel to capture features in a larger range, and convolution branch 3 uses a hole convolution to expand the receptive field to obtain more contextual information.

[0087] Step 5: Select the activation function as ReLU, or Rectified Linear Unit. By outputting negative values ​​as zero and keeping positive values ​​unchanged, nonlinear characteristics are effectively introduced, which enhances the model's expressiveness while reducing the gradient vanishing problem.

[0088] In step 6, a 2×2 maximum pooling operation is performed on the feature maps of each branch obtained in steps 4 and 5 to further extract information of different scales.

[0089] In step 7, the feature maps of each branch in step 6 are concatenated in the depth dimension to form an information map containing multi-scale features. Then, a 1×1 convolution is applied to process the concatenated feature maps to generate a fused multi-scale feature map.

[0090] Step 8: Design skip connections to record high-resolution feature maps, i.e., the first high-resolution feature maps, at each stage of the encoder, and use these feature maps for fusion in the decoder.

[0091] Specifically, the fused multi-scale feature map is obtained by extracting and fusion of the multi-scale convolution branch in the encoder, which is usually the integration of deeper features; the high-resolution feature map is directly extracted from the encoder, which saves the local details and spatial resolution information of the input image. The two are combined through a preset fusion strategy to improve the feature recovery capability. In other words, the high-resolution feature map is the image directly extracted from the encoder.

[0092] Step 9: Save and transfer features. The high-resolution feature map of the encoder stage obtained in step 8 is transferred to the corresponding stage in the decoder through a skip connection.

[0093] Step 10, in the decoder, use a transposed convolution or interpolation method to upsample the feature map obtained in steps 8 and 9, restore it to a high resolution, and obtain a second high-resolution feature map.

[0094] Step 11, fusing the upsampled feature map with the high-resolution feature map of the jump connection. In other words, fusing the first high-resolution feature map with the second high-resolution feature map to obtain a fusion result;

[0095] There are several fusion strategies:

[0096] Concatenation: Concatenate the two feature maps in the depth dimension and then process them through a convolutional layer.

[0097] Specifically, in deep learning, the depth dimension refers to the number of channels in the feature map. In deep learning, images are usually represented as three-dimensional tensors, and their dimensions are in the following order: height, width, and number of channels. The height and width correspond to the spatial size of the image, and the number of channels, i.e., the depth dimension, represents the number of features corresponding to each pixel position in the feature map.

[0098] Weighted fusion: Perform weighted summation on two feature maps, and the weighted coefficient can be obtained through network learning.

[0099] Adaptive fusion: Introduce the attention mechanism to dynamically adjust the fusion weights according to specific tasks.

[0100] Step 12, use 1×1 convolution to map the fused feature map, that is, the fusion result, to the category space to generate the classification result of each pixel, that is, the pixel result;

[0101] Step 13, apply the softmax function to generate the classification probability distribution, assign each pixel to the category with the highest probability, and output the image.

[0102] The assigning of each pixel to the category with the highest probability is specifically achieved as follows:

[0103] First, in semantic segmentation, the last layer of the model usually outputs the class score logits corresponding to each pixel;

[0104] Secondly, the Softmax function is used to convert these score logits into probability distribution. For each pixel position (x, y), the probability that the pixel belongs to category k is Pk(x, y). The result is a vector of length k, and the sum of all probabilities is 1.

[0105] Then, for this probability vector, the "maximum value index" operation is performed to find the category index with the highest probability, and the pixel (x, y) is assigned to this category;

[0106] Finally, the category corresponding to each pixel is output to form the final classification result and output the image.

[0107] Step 14: Use the standard Adam optimization algorithm to train the network and set the hyperparameters of the Adam algorithm: learning rate α, decay rate β of the first-order moment estimate 11 , the decay rate of the second-order moment estimate β2 and the numerical stability constant ε. The specific values ​​are set to: α = 0.001, β 11 =0.9,β2=0.999,ε=10 -8 .

[0108] In step 15, a learning rate decay strategy is adopted, using the StepLR scheduler to reduce the learning rate every certain number of training steps to improve the convergence accuracy when approaching the optimal solution.

[0109] Step 16: Apply data augmentation techniques to the training data. This includes rotation, scaling, flipping, cropping, color conversion, etc. to prevent overfitting during model training.

[0110] Step 17, standardizing the input image, helps speed up training and improve the stability of the model.

[0111] Step 18, the model training process includes forward propagation, loss calculation, back propagation and optimization update. The prediction results of the model are calculated through forward propagation, compared with the true label, and the loss value is calculated. Then, the gradient is calculated through the back propagation algorithm, and the weight of the model is updated using the Adam optimization algorithm.

[0112] Step 19, set the number of training rounds, that is, the number of times the entire data set passes through the network. After each round of training, perform model verification and performance evaluation.

[0113] Step 20: Monitor the training and validation loss values ​​to determine the convergence of the model and whether overfitting occurs.

[0114] Step 21, record and plot the training and validation loss as a function of the number of training rounds.

[0115] Step 22, according to the monitored training indicators, adjust the network's hyperparameters, such as learning rate, batch size, and optimization algorithm hyperparameters. If the loss value is found to be unstable during training, the learning rate can be reduced or the number of training rounds can be increased.

[0116] Step 23: To prevent overfitting, L2 regularization, i.e., weight decay, Dropout, etc., is used to improve the performance of the model on unseen data.

[0117] Step 24, calculate the evaluation index.

[0118] Pixel Accuracy: Calculate the classification accuracy of all pixels, that is, the ratio of correctly classified pixels to the total number of pixels. The formula is:

[0119]

[0120] Intersection over Union, or IoU for short: Calculate the IoU for each category, which is the ratio of the intersection and union of the predicted area and the true area. The mathematical expression is:

[0121]

[0122] Among them, IoU means intersection over union; B pre With B gt They represent the predicted area and the real area respectively; area represents the area.

[0123] Mean Intersection over Union (mIoU): Calculate the average of all categories of IoU for comprehensive evaluation

[0124] The segmentation performance of the model is expressed mathematically as:

[0125]

[0126] Where N is the number of categories, IoU i is the IoU of the i-th category.

[0127] The present invention also provides a sonar image segmentation network implementation system, which can be implemented by executing the process steps of the sonar image segmentation network implementation method, that is, those skilled in the art can understand the sonar image segmentation network implementation method as a preferred implementation mode of the sonar image segmentation network implementation system.

[0128] A sonar image segmentation network implementation system provided by the present invention includes:

[0129] Module M1: Collect sonar images and build training data sets;

[0130] Module M2: preprocessing the training data set, and extracting features of the sonar images of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales;

[0131] Module M3: Maximum pooling of the feature maps of different scales, and splicing them into a spliced ​​information map; convolving the spliced ​​information map to generate a fused multi-scale feature map; recording the fused multi-scale feature map to obtain a first high-resolution feature map;

[0132] Module M4: transpose convolution of the first high-resolution feature map to obtain a second high-resolution feature map; fuse the first high-resolution feature map with the second high-resolution feature map to obtain a fusion result; convolve the fusion result to obtain a pixel result;

[0133] Module M5: distribute the pixel results and output the image.

[0134] Specifically, in the module M1, Gaussian noise is added to the training data set;

[0135] The module M2 includes:

[0136] Module M2.1: Data augmentation of the training dataset;

[0137] Module M2.2: extracting features of the sonar image of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales;

[0138] In the module M2.1, the data enhancement methods include: rotation, translation and scaling;

[0139] In the module M2.2, the features of the sonar image of the training data set are extracted from three scales to obtain feature maps of three different scales; the three scales include: 3×3 convolution kernel, 5×5 convolution kernel and void convolution.

[0140] Specifically, in the module M3, the three feature maps of different scales are pooled by 2×2 maximum pooling and spliced ​​into a spliced ​​information map; the spliced ​​information map is convolved by 1×1 to generate a fused multi-scale feature map;

[0141] In the module M4, the fusion methods include: splicing, weighted fusion and adaptive fusion;

[0142] The weighted coefficients of the weighted fusion are obtained through network learning;

[0143] The adaptive fusion is to adjust the fusion weight according to the task objectives.

[0144] Specifically, in the module M5, a softmax function is applied to generate a classification probability distribution, and then the pixel results are allocated to output an image.

[0145] According to a computer-readable storage medium storing a computer program provided by the present invention, when the computer program is executed by a processor, the steps of the sonar image segmentation network implementation method are implemented.

[0146] An electronic device provided according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the sonar image segmentation network implementation method are implemented.

[0147] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0148] In the description of the present application, it should be understood that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0149] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A sonar image segmentation network implementation method, characterized in that: include: Step S1: Collect sonar images and construct a training data set; Step S2: preprocessing the training data set, and extracting features of the sonar images of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales; Step S3: maximum pooling the feature maps of different scales and splicing them into a spliced ​​information map; convolving the spliced ​​information map to generate a fused multi-scale feature map; recording the high-resolution feature map from the encoder to obtain a first high-resolution feature map; Step S4: transpose convolve the fused multi-scale feature map to obtain a second high-resolution feature map; fuse the first high-resolution feature map with the second high-resolution feature map to obtain a fusion result; convolve the fusion result to obtain a pixel-level prediction result; Step S5: assign the pixel-level prediction result and output an image.

2. The sonar image segmentation network implementation method according to claim 1, characterized in that: In the step S1, Gaussian noise is added to the training data set; In the step S2, it includes: Step S2.1: data augmentation of the training data set; Step S2.2: extracting features of the sonar image of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales; In step S2.1, the data enhancement methods include: rotation, translation and scaling; In the step S2.2, the features of the sonar image of the training data set are extracted from three scales, and feature maps of three different scales are obtained accordingly; the three scales include: 3×3 convolution kernel, 5×5 convolution kernel and void convolution.

3. The sonar image segmentation network implementation method according to claim 2, characterized in that: In the step S3, the three feature maps of different scales are pooled by 2×2 maximum pooling and spliced ​​into a spliced ​​information map; the spliced ​​information map is convolved by 1×1 to generate a fused multi-scale feature map; The high-resolution feature map is an image directly extracted from the encoder; In step S4, the fusion methods include: splicing, weighted fusion and adaptive fusion; The weighted coefficients of the weighted fusion are obtained through network learning; The adaptive fusion is to adjust the fusion weight according to the task objectives.

4. The sonar image segmentation network implementation method according to claim 3 is characterized in that: In the step S5, a softmax function is applied to generate a classification probability distribution, and then the pixel results are allocated to output an image; The steps of allocating include: Step A1: Collect the output of the model to obtain the category score corresponding to each pixel, i.e. logits; Step A2: converting the category scores into probability distributions through a softmax function; Step A3: Based on the probability distribution, the category index with the highest probability is obtained by taking the maximum value index, and then the pixels are assigned to the corresponding categories to form the final classification result and output the image.

5. A sonar image segmentation network implementation system, characterized in that: include: Module M1: Collect sonar images and build training data sets; Module M2: preprocessing the training data set, and extracting features of the sonar images of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales; Module M3: Maximum pooling of the feature maps of different scales, and splicing them into a spliced ​​information map; convolving the spliced ​​information map to generate a fused multi-scale feature map; recording the high-resolution feature map from the encoder to obtain a first high-resolution feature map; Module M4: transpose convolution the fused multi-scale feature map to obtain a second high-resolution feature map; fuse the first high-resolution feature map with the second high-resolution feature map to obtain a fusion result; convolve the fusion result to obtain a pixel-level prediction result; Module M5: distributes the pixel-level prediction results and outputs an image.

6. The sonar image segmentation network implementation system according to claim 5, characterized in that: In the module M1, adding Gaussian noise to the training data set; The module M2 includes: Module M2.1: Data augmentation of the training dataset; Module M2.2: extracting features of the sonar image of the training data set from multiple scales, and correspondingly obtaining feature maps of multiple different scales; In the module M2.1, the data enhancement methods include: rotation, translation and scaling; In the module M2.2, the features of the sonar image of the training data set are extracted from three scales to obtain feature maps of three different scales; the three scales include: 3×3 convolution kernel, 5×5 convolution kernel and void convolution.

7. The sonar image segmentation network implementation system according to claim 6, characterized in that: In the module M3, the three feature maps of different scales are pooled by 2×2 maximum pooling and spliced ​​into a spliced ​​information map; the spliced ​​information map is convolved by 1×1 to generate a fused multi-scale feature map; The high-resolution feature map is an image directly extracted from the encoder; In the module M4, the fusion methods include: splicing, weighted fusion and adaptive fusion; The weighted coefficients of the weighted fusion are obtained through network learning; The adaptive fusion is to adjust the fusion weight according to the task objectives.

8. The sonar image segmentation network implementation system according to claim 7, characterized in that: In the module M5, a softmax function is applied to generate a classification probability distribution, and then the pixel results are allocated and an image is output; The steps of allocating include: Module A1: Collect the output of the model and obtain the category score corresponding to each pixel, i.e. logits; Module A2: convert the category scores into probability distribution through the softmax function; Module A3: Based on the probability distribution, the category index with the highest probability is obtained by taking the maximum value index, and then the pixels are assigned to the corresponding categories to form the final classification result and output the image.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the sonar image segmentation network implementation method according to any one of claims 1 to 4 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the sonar image segmentation network implementation method according to any one of claims 1 to 4 are implemented.

Images