Front-looking sonar image small target recognition method based on SSE-YOLO deep learning model

Abstract

The application discloses a front-looking sonar image small target recognition method based on an SSE-YOLO deep learning model, and belongs to the field of deep learning; the method comprises the following steps: designing a light-weight feature extraction module SSE based on a classic YOLOv3 model, and building a backbone network; fusing a shallower layer feature and a deep layer feature to construct a new prediction layer, and enhancing the small target representation ability; using a convolution attention mechanism module CBAM in a neck network, and strengthening the detection target feature to improve the anti-background interference ability. The method achieves a satisfactory balance between performance and consumption, so that the sonar image target can be detected in real time and quickly under the premise that the device computing capacity is limited. Through the evaluation process, it is proved that the average precision and detection speed of the method are higher than those of the current light-weight mainstream algorithm YOLOv5n, and the embedded deployment can be better realized.

Description

Technical Field

[0001] This invention belongs to the field of target detection in sonar images, specifically involving a method for recognizing small targets in forward-looking sonar images based on the SSE-YOLO deep learning model. Background Technology

[0002] Sonar image-based target detection technology is defined as analyzing and processing sonar images after echo imaging, and then using target detection techniques to achieve detection. It is widely used in civilian and military fields such as underwater topography detection, fish tracking, and torpedo detection. Underwater target detection includes two main methods: traditional machine learning and the recently emerging deep learning based on convolutional neural networks (CNNs).

[0003] Traditional sonar image target detection methods first manually extract contour, texture, and color features from sonar images, then use classifiers to classify and locate targets. Methods such as Support Vector Machines, Singular Value Decomposition, and Independent Component Analysis (ICA) are examples. An inherent flaw of traditional sonar image target detection methods is that they rely on manually extracted image features as the standard for target identification. If the image has few pixels, feature extraction is inaccurate, and the deep features of the sonar image cannot be utilized for decision-making, leading to ineffective learning from the sonar image data and unsatisfactory recognition results. In recent years, deep learning-based target detection methods have been introduced into sonar image target detection. However, most of these methods directly apply mainstream target detection algorithms such as the YOLO series and SSD to sonar image target detection without making adaptive improvements based on practical applications and the characteristics of acoustic images.

[0004] Sonar images have few effective pixels and small imaging areas, making overly complex network architectures unsuitable for feature learning. Furthermore, excessively large network models with deep layers and numerous parameters hinder deployment on performance-constrained embedded devices. Large model size and redundant parameters waste storage resources and slow down detection; excessively deep networks can lead to overfitting of simple sonar image features, affecting detection accuracy and model robustness. Therefore, adaptive improvements are needed for target detection in sonar images using deep learning. Summary of the Invention

[0005] To address the technical problems existing in the background art, this invention provides a method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model. A dual-frequency multi-beam forward-looking sonar is used to acquire acoustic images, with a high-frequency operating mode of 1.2MHz to improve the imaging resolution and quality of the sonar images. Lightweight feature extraction modules PREConv and SSE are constructed, and a novel lightweight backbone network is built to alleviate the overfitting problem caused by overly complex network structures. To address the low detection accuracy of small targets in forward-looking sonar images, a YOLO detection head for detecting even smaller targets is constructed, improving target feature fusion and enabling the semantic information of small targets to flow from deeper to shallower layers, extracting diverse features of the targets. The CBAM attention mechanism is introduced to solve the problems of missed detection and inaccurate localization of small targets during feature extraction.

[0006] To solve the technical problem, the technical solution of the present invention is as follows:

[0007] A method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model, the method comprising:

[0008] S1: Acquire sonar images and construct an underwater small target dataset;

[0009] S2: Construct the initial feature extraction module PREConv and the lightweight feature extraction module SSE, build the backbone feature extraction network, input the dataset into the improved SSE-YOLO model for training, and achieve the feature extraction of the target.

[0010] S3: Improve the neck network, build a YOLO prediction head for detecting smaller targets, and introduce the CBAM attention mechanism to finally obtain the optimized SSE-YOLO deep learning model.

[0011] Furthermore, a multibeam forward-looking sonar was used as the acoustic image acquisition device, and the target to be detected in the acoustic image was labeled using the labeling software LabelImg, thereby constructing an underwater small target dataset.

[0012] Furthermore, the sonar operates at a high frequency of 1.2MHz to improve the imaging resolution of the sonar images. The dataset was labeled using LabelImg software, and the annotated label files were saved in xml format. The dataset contains a total of 17,346 sonar images with three types of labels. The targets include an oxygen cylinder approximately 40cm long and two spheres of different sizes. Each sonar image includes the above three types of targets, with the oxygen cylinder showing different shapes rotated horizontally by 6 degrees. The acquired sonar images were divided into training, validation, and test sets in an 8:1:1 ratio. Data augmentation was performed on the training data using random scaling, flipping, brightness enhancement, contrast enhancement, and Mosaic techniques.

[0013] Furthermore, in step S2, depthwise separable convolutions are used to replace the ordinary convolutions in the initial feature extraction module CBL of the original algorithm network to construct the initial feature extraction module PREConv. The specific structure of the PREConv convolutional block is as follows:

[0014] The PREConv convolutional block includes one depthwise separable convolutional layer: consisting of one channel-wise convolution and one pointwise convolution, with kernel sizes of 3×3 and 1×1 respectively, padding of 1, one batch normalization layer, and one LeakyReLU activation function.

[0015] Furthermore, in step S2, a lightweight feature extraction module SSE is constructed: using the ideas of residual connections, depthwise separable convolution, SE attention mechanism, and proportional channel separation to form a multi-path structure;

[0016] The SSE module performs a 1×1 ordinary convolution operation on the input feature map, initially processing and adjusting the number of channels in the feature map to twice the original number of channels. After the 1×1 ordinary convolution, the feature map is split into two branches; the split ratio is set to 0.25, one branch does not perform any operation on the data, and the other branch, the Dense block, is used for further feature extraction.

[0017] The Dense block maintains the same number of input and output channels. It first separates the input feature map into channels, with each branch having half the number of channels. The right branch remains unchanged, while the left branch consists of two convolutions with a stride of 1: a 1×1 ordinary convolution and a 3×3 depthwise convolution. The two branches are then concatted to add the number of channels and fuse the features. Channel shuffling is used to facilitate information exchange between different groups, ensuring that the channels are fully integrated.

[0018] The SE attention mechanism module is embedded into the SSE module. The SE attention mechanism module obtains the weight factor of each channel feature layer and calculates the importance of each feature channel, thereby making the network model pay more attention to the feature channels related to the target and effectively improving the noise resistance performance of the network model.

[0019] Furthermore, the specific structure of the above three convolutional blocks is as follows:

[0020] A typical convolutional block consists of: one typical convolutional layer with a 1×1 kernel and zero padding, one batch normalization layer, and one ReLU activation function;

[0021] A deep convolutional block consists of one ordinary convolutional layer: a 3×3 kernel with zero padding.

[0022] The attention mechanism SE module includes a global max pooling layer and two fully connected layers. The global max pooling layer is used to transform the input feature map with a height and width of H×W into a 1×1 feature map while keeping the number of channels unchanged. The two fully connected layers better fit the complex correlations between channels and have more non-linearity.

[0023] Furthermore, in step S2, the backbone feature extraction network is constructed from the initial feature extraction module PREConv, the max pooling layer Maxpool, and the SSE module. PREConv is used to extract the initial features of the input feature map. Maxpool is used to reduce the resolution of the feature map, further reduce the amount of computation, reduce the model complexity, and better preserve the texture contour features. The SSE module is used to extract deeper features.

[0024] Furthermore, in step S3, the neck network structure is improved by adopting a multi-scale fusion method based on PANet structure. A bottom-up feature fusion layer is added on the basis of FPN, and the shallow and deep features of the network are fused by upsampling to obtain feature prediction maps of different scales.

[0025] Furthermore, in step S3, the YOLO detection head is improved, and a YOLO prediction layer γ1 that integrates shallower information is constructed, so that the semantic information of small targets can flow from deeper to shallower layers. A new γ1 prediction layer with a scale of 104×104 and a γ2 prediction layer with a scale of 52×52 are constructed, so that the original three prediction layers network becomes a two prediction layers network.

[0026] Furthermore, in step S3, CBAM is added to the multi-scale detection layer to improve the model's ability to represent target features with almost no increase in model size, thereby improving the model's detection performance. Compared with the prior art, the advantages of this invention are:

[0027] (1) The construction of the new YOLO prediction head and the improvement of the feature fusion network have integrated deeper semantic information and shallower positional information, which has improved the detection accuracy of small targets in sonar images.

[0028] (2) The construction of lightweight feature extraction modules PREConv and SSE, and the construction of a brand-new backbone network, reduced the model size and the number of network parameters, thus achieving network lightweighting.

[0029] (3) In terms of detection speed, both the algorithm and the algorithm mentioned above meet the real-time requirement of FPS≥30. However, the algorithm is more balanced, better balancing model size, detection accuracy and detection speed, and is more suitable for application in micro-sized underwater embedded devices. Attached Figure Description

[0030] Figure 1 This is a flowchart of the small target detection method based on sonar images described in this invention;

[0031] Figure 2 This is a schematic diagram of target annotation in a sonar image provided by the present invention;

[0032] Figure 3 A schematic diagram of the network structure of the improved YOLO backbone feature extraction network provided by this invention;

[0033] Figure 4 This is a network structure diagram of the feature extraction module SSE described in this invention;

[0034] Figure 5 This is a schematic diagram of the improved feature fusion network structure described in this invention. Detailed Implementation

[0035] The specific implementation of the present invention is described below with reference to embodiments:

[0036] It should be noted that the structures, proportions, sizes, etc. shown in this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0037] Furthermore, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity of description and are not intended to limit the scope of the invention. Any changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0038] Example 1:

[0039] Current image recognition methods for small underwater targets almost universally suffer from low recognition rates. The root cause lies in the small size and weak signal strength of these targets, coupled with severe environmental noise interference, resulting in unclear textures, blurred contours, and motion blur in the sonar image, thus increasing the difficulty of recognition. Traditional recognition methods rely on manual feature extraction, meaning that image quality directly impacts recognition performance. While deep learning-based methods offer superior performance, they rarely adapt to the imaging characteristics of sonar images. Poor image quality and low pixel count significantly hinder feature extraction by convolutional networks. Therefore, regardless of the method employed, the key challenge in recognizing small targets in sonar images lies in improving image quality and extracting more effective information to aid target recognition, thereby increasing the recognition rate.

[0040] Combination Figure 1-5 This invention proposes an underwater small target identification method based on SSE-YOLO, the method comprising the following steps:

[0041] S1: Image Acquisition and Dataset Construction. The dataset for this invention was acquired in a university's comprehensive experimental water tank. A dual-frequency multibeam forward-looking sonar was used as the acoustic image acquisition device. The sonar operated at a high frequency of 1.2MHz, which greatly improved the imaging resolution of the sonar images. The labeling software LabelImg was used to label the targets to be detected in the acoustic images, thereby constructing an underwater small target dataset.

[0042] S2: Improved Backbone Feature Extraction Network. A lightweight feature extraction module (SSE) is designed, and a new backbone network (SSE) is built with a max-pooling layer (MaxPool) to replace the original YOLOv3 backbone network Darknet53. Sonar images are input into SSE-YOLO for training to achieve target feature extraction.

[0043] S3: Improved Neck Network and Prediction Head. The neck network employs a Path Aggregation Network (PAN) to construct a YOLO prediction head for detecting smaller targets, achieving feature fusion and extracting diverse target features. A CBAM attention mechanism is introduced to address the issues of missed detections and inaccurate localization of small targets during feature extraction.

[0044] Step S1, which involves acquiring sonar images and constructing a dataset, specifically includes the following steps:

[0045] S1.1: The dataset was labeled using LabelImg software, and the annotated label files were saved in .xml format. The dataset contains 17,346 sonar images with three label classes: "bottle", "ball1", and "ball2". Figure 2 The central target corresponds to an oxygen cylinder approximately 40cm long, while the left and right targets are two spheres of different sizes. Each sonar image includes all three types of targets, with the oxygen cylinder exhibiting different configurations that rotate horizontally in 6-degree increments.

[0046] S1.2: In this embodiment, the acquired images are divided into training, validation, and test sets in an 8:1:1 ratio. During training, data augmentation techniques such as random scaling, flipping, brightness enhancement, contrast enhancement, and Mosaic are used to enrich the dataset and improve the robustness of the model.

[0047] Combination Figure 3 Step S2 involves constructing the SSE-YOLO backbone network. Specifically, it includes the following steps:

[0048] S2.1: Improved Initial Feature Extraction Module. This invention replaces the ordinary convolution in the original algorithm's initial feature extraction module CBL with depthwise separable convolutions to construct the initial feature extraction module PREConv. Compared with the original convolutional blocks, depthwise separable convolutions reduce the network model size, the number of algorithm parameters, and the computational cost. The specific structure of the two convolutional blocks is as follows:

[0049] CBL convolutional block: includes 1 regular convolutional layer (3×3 kernel, padding 1), 1 batch normalization layer, and 1 LeakyReLU activation function;

[0050] PREConv convolutional block: includes 1 depthwise separable convolutional layer (divided into 1 channel-wise convolution and 1 pointwise convolution, with kernel sizes of 3×3 and 1×1 respectively, and padding of 1), 1 batch normalization layer, and 1 LeakyReLU activation function.

[0051] S2.2: Constructing a lightweight feature extraction module SSE. This invention uses residual connections, depthwise separable convolutions, SE attention mechanisms, and proportional channel separation to form a multi-path structure, which horizontally increases the network width and expands the receptive field, enhancing the network's coverage and sensitivity for small target feature extraction.

[0052] This invention constructs a feature extraction module SSE, such as Figure 4 (a) The SSE module performs a 1×1 ordinary convolution operation on the input feature map, initially processing and adjusting the number of channels in the feature map to twice the original number. After the 1×1 ordinary convolution, the feature map is split into two branches. This invention sets the split ratio to 0.25. One branch does not perform any operation on the data, while the other branch, the Dense block, is used for further feature extraction.

[0053] Dense blocks maintain the same number of input and output channels. They first perform channel separation on the input feature map, with each branch having half the number of channels. Figure 4 (b) The right branch remains unchanged, while the left branch consists of two convolutions with a stride of 1: a 1×1 ordinary convolution and a 3×3 depthwise convolution. A concat operation is performed on the two branches, adding the number of channels to fuse features. Channel shuffling is then used to facilitate information exchange between different groups, ensuring thorough channel fusion.

[0054] The SE attention mechanism module is embedded into the SSE module. The SE attention mechanism module calculates the importance of each feature channel by obtaining the weight factors of each channel's feature layer, thereby allowing the network model to focus more on feature channels relevant to the target, effectively improving the network model's noise resistance. The specific structure of the three convolutional blocks is as follows:

[0055] A typical convolutional block consists of one typical convolutional layer (1×1 kernel, padded with 0), one batch normalization layer, and one ReLU activation function.

[0056] Depth convolution block: consists of one depth convolutional layer (3×3 kernel, padded with 0);

[0057] The attention mechanism (SE module) consists of a global max-pooling layer and two fully connected layers. The global max-pooling layer transforms the input feature map (H×W) into a 1×1 feature map while maintaining the same number of channels. The two fully connected layers better fit the complex relationships between channels and provide more non-linearity.

[0058] S2.3: Constructing a novel lightweight backbone network. In this invention, the backbone network is constructed from an initial feature extraction module PREConv, a max-pooling layer Maxpool, and an SSE module. PREConv is used to extract initial features from the input feature map; Maxpool is used to reduce the resolution of the feature map, further reducing computation and model complexity, while better preserving texture contour features. The SSE module is used to extract deeper features.

[0059] Step S3, which improves the neck network and predicts the head, specifically includes the following steps:

[0060] S3.1: Improve the neck network structure. Small sonar targets have limited pixel information and lack texture features, and relevant information is easily lost during convolutional neural network feature extraction. In the process of extracting features from input images through convolution, shallow feature maps have high resolution, small receptive fields, and rich spatial information. As the network deepens, the resolution of deeper feature maps continuously decreases, while the receptive field increases, enhancing semantic information.

[0061] Based on the high-resolution representation information of shallow feature maps and the high semantic information of deep feature maps, this invention adopts a multi-scale fusion method with a PAN structure. It adds a bottom-up feature fusion layer on the basis of FPN and uses an upsampling method to fuse the shallow and deep features of the network to obtain feature prediction maps at different scales.

[0062] S3.2: Improved YOLO detection head. For small target detection, as the feature map resolution decreases, the number of pixels occupied by small target features in the feature map also gradually decreases. For example, for a small target with a size of 13×13, after multiple pooling downsampling, the effective feature information of the small target is less, and it is easily identified by the model as background noise and missed.

[0063] Compared to the three prediction layers of YOLOv3 (scales of 52×52, 26×26, and 13×13 respectively), this invention constructs a YOLO prediction layer γ1 that integrates shallower information, enabling the semantic information of small targets to flow from deeper to shallower layers. For example... Figure 5 As shown, the shallow network L4 in the SSE-YOLO backbone network is fused with the feature map of the same scale output by the feature layer D2 in the neck network after two upsamplings to construct a new γ1 prediction layer with a scale of 104×104. The feature map output by the feature layer D1 in the neck network is fused with the feature map output by the deeper layer D5 to construct a γ2 prediction layer with a scale of 52×52. This reduces the original three prediction layers to two prediction layers, thereby reducing the computational cost and parameters of the network and enhancing the network's ability to detect small objects.

[0064] S3.3: This invention improves the model's ability to express target features and enhances the model's detection performance by adding CBAM to the multi-scale detection layer without increasing the model size.

[0065] Example 2:

[0066] The evaluation process of this invention is described below:

[0067] In this embodiment, the obtained dataset is divided into training, validation, and test sets in an 8:1:1 ratio, and the trained model is used to perform detection on the test set. The evaluation metrics for object detection mainly include the mean accuracy (mAP) of the test samples, the number of model parameters, the model size, and the detection speed (FPS). The specific formula for calculating mAP is as follows:

[0068]

[0069]

[0070]

[0071]

[0072] In this context, TP represents the number of positive samples identified as positive samples, FN represents the number of positive samples misclassified as negative samples, FP represents the number of negative samples misclassified as positive samples, n represents the number of classes in the sample, and P, R, and AP are defined as precision, recall, and single-class precision, respectively.

[0073] The test used a classic lightweight convolutional neural network algorithm to identify sonar images, and compared it with the method described in this invention. The results are shown in Table 1 below.

[0074] Table 1

[0075]

[0076] As can be seen from the table, the present invention has higher accuracy and detection performance for small targets in sonar images. With a model size of only 5.2M and a parameter count of only 2.48M, its mAP value reaches 88.1%, which is better than other lightweight networks such as YOLOv5n.

[0077] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

[0078] Many other changes and modifications can be made without departing from the concept and scope of this invention. It should be understood that this invention is not limited to the specific embodiments, and the scope of this invention is defined by the appended claims.

Claims

1. A method for small target recognition in forward-looking sonar images based on an SSE-YOLO deep learning model, characterized in that, The method includes: S1: Acquire sonar images and construct an underwater small target dataset; S2: Construct the initial feature extraction module PREConv and the lightweight feature extraction module SSE, build the backbone feature extraction network, input the dataset into the improved SSE-YOLO model for training, and achieve the feature extraction of the target. S3: Improve the neck network, build a YOLO prediction head for detecting smaller targets, and introduce the CBAM attention mechanism to finally obtain the optimized SSE-YOLO deep learning model; In step S2, depthwise separable convolutions are used to replace the ordinary convolutions in the original algorithm's initial feature extraction module CBL, constructing the initial feature extraction module PREConv. The specific structure of the PREConv convolutional block is as follows: PREConv convolutional block includes one depthwise separable convolutional layer: consisting of one channel-wise convolution and one pointwise convolution, with kernel sizes of 3×3 and 1×1 respectively, padding of 1, one batch normalization layer, and one LeakyReLU activation function; In step S2, a lightweight feature extraction module SSE is constructed: using residual connections, depthwise separable convolutions, SE attention mechanisms, and proportional channel separation to form a multi-path structure; The SSE module performs a 1×1 ordinary convolution operation on the input feature map, initially processing and adjusting the number of channels in the feature map to twice the original number of channels. After the 1×1 ordinary convolution, the feature map is split into two branches; the split ratio is set to 0.25, one branch does not perform any operation on the data, and the other branch, the Dense block, is used for further feature extraction. The Dense block maintains the same number of input and output channels. It first separates the input feature map into channels, with each branch having half the number of channels. The right branch remains unchanged, while the left branch consists of two convolutions with a stride of 1: a 1×1 ordinary convolution and a 3×3 depthwise convolution. The two branches are then concatted, adding the number of channels and fusing the features. Channel shuffling is used to facilitate information exchange between different groups, ensuring full fusion of channels. The SE attention mechanism module is embedded into the SSE module. The SE attention mechanism module obtains the weight factor of each channel feature layer and calculates the importance of each feature channel, thereby making the network model pay more attention to the feature channels related to the target and effectively improving the noise resistance performance of the network model.

2. The method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model according to claim 1, characterized in that, A multibeam forward-looking sonar was used as the acoustic image acquisition device, and the target to be detected in the acoustic image was labeled using the labeling software LabelImg, thereby constructing an underwater small target dataset.

3. The method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model according to claim 2, characterized in that, The sonar operates at a high frequency of 1.2MHz to improve the imaging resolution of the sonar images. The dataset was labeled using LabelImg software, and the annotated label files were saved in XML format. The dataset contains a total of 17,346 sonar images with three types of labels. The targets include an oxygen cylinder approximately 40cm long and two spheres of different sizes. Each sonar image includes the above three types of targets, with the oxygen cylinder showing different shapes rotated horizontally by 6 degrees. The acquired sonar images were divided into training, validation, and test sets in an 8:1:1 ratio. Data augmentation was performed on the training data using random scaling, flipping, brightness enhancement, contrast enhancement, and Mosaic techniques.

4. The method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model according to claim 1, characterized in that, The specific structure of the above three convolutional blocks is as follows: A typical convolutional block consists of: one typical convolutional layer with a 1×1 kernel and zero padding, one batch normalization layer, and one ReLU activation function; A deep convolutional block consists of one ordinary convolutional layer: a 3×3 kernel with zero padding. The attention mechanism SE module includes a global max pooling layer and two fully connected layers. The global max pooling layer is used to transform the input feature map with a height and width of H×W into a 1×1 feature map while keeping the number of channels unchanged. The two fully connected layers better fit the complex correlations between channels and have more non-linearity.

5. The method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model according to claim 4, characterized in that, In step S2, the backbone feature extraction network is constructed from the initial feature extraction module PREConv, the max pooling layer Maxpool, and the SSE module. PREConv is used to extract the initial features of the input feature map. Maxpool is used to reduce the resolution of the feature map, further reduce the amount of computation, reduce the model complexity, and better preserve the texture contour features. The SSE module is used to extract deeper features.

6. The method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model according to claim 1, characterized in that, In step S3, the neck network structure is improved by adopting a multi-scale fusion method based on PANet structure. A bottom-up feature fusion layer is added on the basis of FPN, and the shallow and deep features of the network are fused by upsampling to obtain feature prediction maps at different scales.

7. The method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model according to claim 1, characterized in that, In step S3, the YOLO detection head is improved, and a YOLO prediction layer γ1 that integrates shallower information is constructed, so that the semantic information of small targets can flow from deeper to shallower layers. A new γ1 prediction layer with a scale of 104×104 and a γ2 prediction layer with a scale of 52×52 are constructed, so that the original three prediction layers network becomes a two prediction layers network.

8. The method for small target recognition in forward-looking sonar images based on the SSE-YOLO deep learning model according to claim 1, characterized in that, In step S3, CBAM is added to the multi-scale detection layer to improve the model's ability to express target features with almost no increase in model size, thereby improving the model's detection performance.

Images