Method and system for detecting underwater small targets using synthetic aperture sonar images

Abstract

The present invention discloses a method and system for detecting underwater small targets in synthetic aperture sonar images. The method includes processing acoustic echo data received by an array element, and obtaining image slices that meet size requirements through redundant cutting algorithm processing; inputting image slices one by one into a pre-established and trained small target detection model to obtain target detection results for each slice, and then using secondary non-maximum suppression to screen repeated recognition results of the same target in the slice, and finally realizing underwater small target detection in synthetic aperture sonar images; the small target detection model uses an improved SSD network to realize lightweight target detection, and the improved SSD network includes a basic feature extraction network, and the basic feature extraction network uses an improved feature extraction module. The present invention can detect underwater small targets, effectively improve the accuracy of target detection, and avoid the problem of target missed detection, providing an effective solution for the task of underwater small target detection in synthetic aperture sonar images.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic signal processing, and in particular to a method and system for detecting underwater small targets using synthetic aperture sonar images. Background Art

[0002] Synthetic Aperture Sonar (SAS) obtains seafloor sonar images by coherently processing multiple echo signals from the array and synthesizing an equivalent virtual long aperture. Compared with side-scan sonar images, the most significant advantage of SAS images is that they have higher azimuth resolution, and the theoretical resolution is independent of the target distance and the acoustic wave frequency band used. However, the actual physical size of the small target of interest is small, resulting in fewer effective pixels and a lower relative pixel ratio in the SAS image, which poses a challenge to subsequent target detection and recognition.

[0003] Traditional SAS image small target detection first extracts the target "slice" of interest (Regions of Interest, ROI) from the SAS image; then, the target in the slice is segmented, and then the features of the segmented image are further extracted; finally, the extracted features are classified and identified using a classifier. However, the slice selection algorithm has poor robustness, the design of the target segmentation and feature extraction algorithm is very difficult and easy to cause the loss of useful information, and the shallow classifier is also easy to reach the performance bottleneck. Artificial intelligence algorithms with deep learning as the core have attracted great attention from all over the world. It has fundamentally changed the way people deal with complex and uncertain problems. Compared with traditional machine learning and signal processing methods, deep learning simulates the hierarchical system of the human visual nervous system and contains more hidden unit layers. Through the nonlinear transformation of the original data layer by layer, a higher-level and more abstract feature expression can be obtained. The high-level expression can enhance the distinguishing ability of the input data and weaken the adverse effects of irrelevant factors. Therefore, it is of great theoretical and practical significance to improve the detection and recognition accuracy of small targets of interest in SAS images by drawing on deep learning technology. At present, the target detection model based on convolutional neural networks (CNN) realizes the unification of target detection and recognition in one framework. Object detection models can be divided into two categories: candidate region-based and regression-based. The candidate region-based models are mainly represented by R-CNN (Region-CNN), Fast R-CNN (Fast Region-CNN) and Faster R-CNN (Faster Region-CNN). The regression-based models are mainly represented by YOLO (You Only Look Once) and SSD (Single Shot Detection). The SSD model takes into account both detection accuracy and detection speed. However, the input image size of both types of object detection models is relatively small. For example, the input size of the original image of Faster R-CNN based on the candidate region model is 1000 pixels × 600 pixels, the input size of the original image of SSD is 300 pixels × 300 pixels or 512 pixels × 512 pixels; the input size of the original image of YOLO is 416 pixels × 416 pixels or 544 pixels × 544 pixels. Therefore, a simple downsampling operation cannot guarantee the detection effect of the above object detection models on small objects of interest.

[0004] In summary, there is an urgent need for a detection method for underwater small targets suitable for high-resolution and large-size SAS images to avoid missed detection and improve target detection accuracy. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a method and system for detecting underwater small targets using synthetic aperture sonar images.

[0006] In order to achieve the above object, the present invention proposes a method for detecting underwater small targets using synthetic aperture sonar images, the method comprising:

[0007] Step 1) processing the acoustic echo data received by the array element to obtain a synthetic aperture sonar image, and obtaining a number of image slices that meet the size requirements through a redundant cutting algorithm;

[0008] Step 2) input the image slices one by one into the pre-established and trained small target detection model to obtain the target detection result of each slice, and then use the secondary non-maximum suppression to screen the repeated recognition results of the same target in the slice, and finally realize the underwater small target detection of synthetic aperture sonar image;

[0009] The small target detection model adopts an improved SSD network to achieve lightweight target detection. The improved SSD network includes a basic feature extraction network, and the basic feature extraction network adopts an improved feature extraction module.

[0010] Preferably, the step 1) specifically includes:

[0011] The acoustic echo data received by the array element is processed to obtain a synthetic aperture sonar image I with a size of N×N;

[0012] After being processed by the redundant cutting algorithm, several image slices I are obtained. r,c,h,s , where r represents the column coordinate of the image slice in I, c represents the row coordinate of the image slice in I, h represents the size of the image slice, s represents the cutting step of the image slice, adjacent image slices overlap in the horizontal and vertical directions, the single-sided overlap ratio is a set value to ensure that the small target of interest exists completely in the image slice, the four vertex images of I overlap on two sides, the edge image overlaps on three sides, and the remaining image slices overlap on four sides.

[0013] Preferably, the small target detection model includes a basic feature extraction network, an additional feature extraction network, a candidate box generation module and a detection result prediction module connected in sequence, wherein:

[0014] The basic feature extraction network includes two improved feature extraction modules, which output feature layers of two sizes respectively;

[0015] The additional feature extraction network includes 4 convolutional layers, which respectively output feature layers of 4 sizes;

[0016] The candidate box generation module is used to extract candidate boxes of different numbers and sizes from the feature layers of the above six sizes according to predefined scales and aspect ratios;

[0017] The detection result prediction module is used to determine the type and position of the target in the candidate frame, and optimize the target in the candidate frame using a non-maximum suppression algorithm.

[0018] Preferably, the input of the improved feature extraction module is feature D, and the output is for:

[0019]

[0020] Wherein, D' is the output of the first branch, the first branch is a sequentially connected expansion layer, a segmentation layer, three parallel branches, a splicing layer, a channel selection component and a compression layer, the expansion layer and the compression layer both include a BN layer and a ReLU layer, the parallel branch includes a sequentially connected multi-scale component and a space selection component, wherein,

[0021] The expansion layer is responsible for expanding the input feature channel;

[0022] The segmentation layer is responsible for channel cutting;

[0023] The multi-scale component is responsible for multi-scale decomposition of input information;

[0024] The spatial selection component is responsible for selecting useful spatial information in the multi-scale layer;

[0025] The splicing layer is responsible for channel integration;

[0026] The channel selection component is used to select channels containing important information by learning weights;

[0027] The compression layer is responsible for compressing the feature channels into a number consistent with the input features.

[0028] Preferably, the processing process of the feature extraction module specifically includes:

[0029] The input feature D passes through the expansion layer, and its output feature D ex for:

[0030] D ex =F ex (D),D∈Φ H×H×M

[0031] Among them, F ex This is a channel expansion operation. The convolution kernel size of the expansion layer is 1×1, and the number of output channels is the number of channels after expansion M. + ;

[0032] D exEnter the segmentation layer and the multi-scale components of each branch. The multi-scale component of the kth branch outputs D k branchs for:

[0033]

[0034] in, The features after segmentation are evenly segmented into three parallel branches; The output feature map size is H×H and the number of channels is M + / 3,F branchs It is a multi-scale depth-separable dilated convolution operation.

[0035] D k branchs Input space selection component, whose output features Satisfy the following formula:

[0036]

[0037]

[0038] in, The output feature map size is H×H and the number of channels is M + / 3; is the spatial weight of the kth branch, σ is the activation function, f 3*3 is a convolution operation with a convolution kernel size of 3×3, F concat is the channel splicing operation, AvgPool is the average pooling operation, and MaxPool is the maximum pooling operation;

[0039] Use the splicing layer to 3 Concatenate in the channel dimension, the concatenated output feature D concat for:

[0040]

[0041] Among them, D concat The size of the output feature map is H×H, and the number of channels is M + , F concat It is a channel splicing operation;

[0042] D concat Input channel selection component, whose output feature D chanel for:

[0043] s chanel =MLP(AvgPool(D concat ))+MLP(MaxPool(D concat ))

[0044] Dchanel =s chanel ·D concat

[0045] Among them, D chanel The size of the output feature map is H×H, and the number of channels is M + ,s chanel D concat The channel weight coefficient, MLP is a shared fully connected network;

[0046] Right D chanel Channel compression is performed, and the characteristic D' after channel compression is:

[0047] D'=F sq (D chanel ),D'∈Φ H×H×M

[0048] Among them, the size of the output feature map of D' is H×H, and the number of channels is equal to the number of channels of the input feature D, which is M, F sq For channel compression operation, the convolution kernel size is 1×1.

[0049] Preferably, the processing process of the detection result prediction module specifically includes:

[0050] For the detection information of the image slice B r,c,i,j , the overlapping suggestion boxes are removed by the secondary non-maximum suppression algorithm Soft-NMS, and finally the bounding box with the highest score after regression correction at each position in each category is obtained, which is the small target detection result

[0051]

[0052] Where r represents the column coordinate in the synthetic aperture sonar image I, c represents the row coordinate in I, and i represents the type of target; i is the candidate box with the highest score in the current cycle, b r,c,i is the detection information at the current position, iou represents the intersection-over-union function, T t is the IoU threshold.

[0053] Preferably, the method further comprises a small target detection model training step, specifically comprising:

[0054] Collect raw sonar data from real underwater environments, process them to obtain synthetic aperture sonar images of standard size, and annotate them to generate training sets;

[0055] The training set data is input into the improved SSD network in sequence until the training requirements are met, and a trained small target detection model is obtained.

[0056] On the other hand, the present invention provides a synthetic aperture sonar image underwater small target detection system, the system comprising:

[0057] A processing module is used to process the acoustic echo data received by the array element to obtain a synthetic aperture sonar image, and obtain a number of image slices that meet the size requirements through a redundant cutting algorithm;

[0058] A platform deployment module for deploying pre-built and trained small object detection models to embedded platforms; and

[0059] The detection output module is used to input the image slices one by one into the pre-established and trained small target detection model to obtain the target detection results of each slice, and then use the secondary non-maximum suppression to screen the repeated recognition results of the same target in the slice, and finally realize the underwater small target detection of synthetic aperture sonar images;

[0060] The small target detection model adopts an improved SSD network to achieve lightweight target detection. The improved SSD network includes a basic feature extraction network, and the basic feature extraction network adopts an improved feature extraction module.

[0061] Compared with the prior art, the advantages of the present invention are:

[0062] The present invention combines synthetic aperture sonar with an improved SSD target detection model and proposes an underwater small target detection model. This method solves the problems of missed detection and low underwater small target detection accuracy in the existing methods in an end-to-end manner. By improving the feature extraction unit, the adaptability of the network to underwater small targets is improved. In addition, the integrity of underwater small target detection is guaranteed by adopting redundant cutting. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The invention provides an underwater small target detection method and system implementation framework based on improved SSD;

[0064] Figure 2 is an improved feature extraction module used in the present invention;

[0065] Figure 3 It is an improved SSD network structure used in the present invention;

[0066] Figure 4 It is an improved HRSSD underwater small target detection method used in the present invention;

[0067] Figure 5 This is the underwater small target detection effect of SAS image provided by the present invention Figure 1 ;

[0068] Figure 6This is the underwater small target detection effect of SAS image provided by the present invention Figure 2 ;

[0069] Figure 7 This is the underwater small target detection effect of SAS image provided by the present invention Figure 3 . DETAILED DESCRIPTION

[0070] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0071] Example 1

[0072] Embodiment 1 of the present invention proposes a method for detecting small underwater targets using synthetic aperture sonar images.

[0073] The present invention includes synthetic aperture sonar image preprocessing, data set production, model training and platform deployment. The first step is to use synthetic aperture sonar to collect seabed target images, annotate the data and generate a target detection data set; the second step is to initialize the training parameters, train the improved SSD model, and perform quality assessment on the target detection results; the third step is to deploy the trained target detection model to the edge computing platform to achieve online detection of underwater small targets and result output. The overall process flow chart is as follows Figure 1 As shown, the specific steps are as follows:

[0074] Step 1: Preparation of underwater target detection dataset

[0075] Step 1-1, using the synthetic aperture sonar acquisition submodule to collect raw sonar data from a real underwater environment;

[0076] Step 1-2: Use the open source software labelImg to label the sonar images;

[0077] Step 1-3: According to the VOC dataset format, the labeled synthetic aperture sonar images are divided into a training sample set and a test sample set using the principle of random division.

[0078] Step 2: Model training

[0079] Step 2-1: Build the environment required for the training platform on the deep learning server, including open source software Anaconda, Pytorch, and Torchvision, and set the model training initialization parameters, including batchsize and epoch;

[0080] Step 2-2: Build an improved feature extraction module. The improved feature extraction module (Inverted ResidualBlock Plus and Plus, IRB ++ )like Figure 2 As shown. IRB ++ The module selects input information from three dimensions: multi-scale, spatial, and channel. The improved attention component is denoted as BSC (Branch Spatial Channel Attention). IRB ++ Modules use IRB + The module adopts the strategy of "expanding" first and then "compressing", and is composed of an expansion layer, a segmentation layer, a multi-scale (Selective Dilated Kernel, SDK) component, a spatial selection (Spatial Attention, SA) component, a splicing layer, a channel selection (Channel Attention, CA) component and a compression layer. The expansion layer is responsible for the expansion of the input feature channel; the segmentation layer is responsible for channel cutting; the multi-scale component is responsible for the multi-scale decomposition of the input information; the spatial selection component is responsible for selecting useful spatial information in the multi-scale layer; the splicing layer is responsible for channel integration; the channel selection component selects channels containing important information by learning weights; the compression layer is responsible for compressing the feature channels into a number consistent with the input features.

[0081] For an arbitrary input feature D∈Φ H×H×M , where H×H is the size of the input feature and M is the number of channels of the input feature. Input feature D enters IRB ++ The two branch networks of the module: the left branch is responsible for the feature extraction and selection of the small target of interest; the right branch keeps the input feature D unchanged and finally adds it to the output feature of the upper branch network. For the lower branch network, the input feature D first passes through the expansion layer, and the mathematical expression of its output feature is:

[0082] D ex =F ex (D),D∈Φ H×H×M (1)

[0083] Where D is the original input feature, D ex is the feature after the expansion layer, F ex This is a channel expansion operation. The convolution kernel size of the expansion layer is 1×1 and the number of output channels is M. + .

[0084] Next, output feature D ex Submit to IRB ++ The mathematical expression of the output features of the segmentation layer and multi-scale component is:

[0085]

[0086] in is the feature after segmentation, k represents the kth branch, there are 3 branches in total, and it is evenly segmented by channel; It is the feature map output by the multi-scale component, the feature map size is H×H, and the number of channels is M + / 3,F branchs It is a multi-scale depth-separable dilated convolution operation. In the multi-scale component, the relationship between the actual coverage of the convolution kernel and the dilation rate is as follows:

[0087] R rdf =R+(κ-1)(R-1) (3)

[0088] Where R rdf is the actual coverage of the convolution kernel of this layer, R is the size of the convolution kernel of this layer, and κ is the dilation rate of this layer. For example, for a dilated convolution layer with a convolution kernel size of 3×3 and a dilation rate of 2, the actual coverage of the convolution kernel is R rdf =5. Increasing the hole rate κ can expand the receptive field of the convolutional layer. That is, reducing the computational cost of the model and reducing the number of model parameters. In this paper, κ = 1, 2, that is, R rdf =3,5.

[0089] Then, the output features The mathematical expression of the output feature is sent to the spatial selection component:

[0090]

[0091]

[0092] in is the spatial weight of the kth branch, σ is the activation function, f 3*3 is a convolution operation with a convolution kernel size of 3×3, F concat is the channel splicing operation, AvgPool is the average pooling operation, and MaxPool is the maximum pooling operation. The output of the spatial selection component has an output feature map size of H×H and a channel number of M + / 3.

[0093] Then, the concatenation layer is used to When concatenated in the channel dimension, the mathematical expression of the output feature is:

[0094]

[0095] Where D concat is the output feature after concatenation, the size of the output feature map is H×H, and the number of channels is M + Equal to the number of channels after expansion, F concat This is a channel splicing operation.

[0096] Next, D concatThe mathematical expression of the output feature of the channel selection component is:

[0097] s chanel =MLP(AvgPool(D concat ))+MLP(MaxPool(D concat )) (7)

[0098] D chanel =s chanel ·D concat (8)

[0099] Where D chanel The output features of the channel selection component, the size of the output feature map is H×H, and the number of channels is M + ,s chanel D concat The channel weight coefficient, MLP shares a fully connected network, AvgPool is an average pooling operation, and MaxPool is a maximum pooling operation.

[0100] Then, for D chanel For channel compression, the mathematical expression is:

[0101] D'=F sq (D chanel ),D'∈Φ H×H×M (9)

[0102] Where D' is the feature after channel compression, the size of the feature map is H×H, the number of channels is M, which is equal to the number of channels of the input feature D, and F sq For channel compression operation, the convolution kernel size is 1×1.

[0103] Through the above calculation, we can finally get IRB ++ The output characteristic mathematical expression of the module is:

[0104]

[0105] in For IRB ++ The output features of the module, D∈Φ H×W×M , the feature map size is H×H, and the number of channels is M.

[0106] Step 2-3: Build the improved target detection SSD-MV3P. The network structure is as follows Figure 3 In the SSD-MV3 model, IRB ++ Replacement of IRB +Module, generate a new lightweight target detection network, denoted as SSD-MV3P. The network structure of SSD-MV3P is consistent with that of SSD-MV3, including four parts: basic feature extraction network, additional feature extraction network, candidate box generation and detection result prediction. The basic network of SSD-MV3P consists of the first 14 convolutional layers, and the additional feature extraction network consists of the last 4 convolutional layers. The input channel, intermediate channel, output channel, step size and activation function (RE represents ReLU activation function, HS represents H-Swich activation function) of each convolutional layer are consistent with SSD-MV3; the difference is that SSD-MV3P uses IRB ++ Replace the IRB in SSD-MV3 + module, IRB ++ The number of branches corresponding to the module is 2, and the convolution kernel size is 3×3 and 5×5. Out_Conv1 and Out_Conv2 in the basic network, and the additional feature extraction networks Conv19_1, Conv19_2, Conv19_3 and Conv19_4 are used to generate candidate boxes, where the input feature map sizes of Out_Conv1, Out_Conv2, Conv19_1, Conv19_2, Conv19_3 and Conv19_4 are 19×19, 10×10, 5×5, 3×3, 2×2 and 1×1. The candidate box generation part extracts candidate boxes of different numbers and sizes from the feature layers of the above 6 scales according to the pre-defined scales and aspect ratios; the convolution prediction part judges the type and position of the object in the candidate box, and optimizes the object in the candidate box using the non-maximum suppression algorithm.

[0107] Step 2-4: Monitor the training process and test results of the improved SSD network in real time, and stop training when the evaluation indicators meet the requirements.

[0108] Step 3: Platform deployment

[0109] Step 3-1: Build the target detection operating environment on the edge computing platform.

[0110] Step 3-2: Deploy the trained target detection model to the platform, build a high-resolution large-size SAS image underwater small target detection method HRSSD, and the method flow is as follows: Figure 4As shown. SAS images have high resolution and low relative pixel ratio of small targets of interest. Directly inputting the original SAS images into SSD-MV3 cannot guarantee the detection effect of small targets of interest. HRSSD adopts the strategy of cutting first, detecting later, and then confirming to detect and identify small targets of interest in SAS images. HRSSD includes the following three steps: first, redundant cutting of high-resolution and large-size SAS images is performed row by row and column by column; then, the improved small target detection model SSD_MV3P is used to detect and identify the image slices respectively; finally, Soft-NMS is used to perform secondary non-maximum suppression on the detection results, and the final results are displayed on high-resolution and large-size SAS images.

[0111] Assuming that the size of the high-resolution large-size SAS image I is N×N, the mathematical description of the image slices processed by the cutting algorithm is as follows:

[0112]

[0113] Among them I r,c,h,s Represents image slices, r represents the column coordinates in SAS image I, c represents the row coordinates in SAS image I, h represents the size of the image slice, and s represents the cutting step of the image slice. In order to ensure the integrity of the small target of interest in the image slice, adjacent image slices overlap in the horizontal and vertical directions, and the single-side overlap ratio is 20% (to ensure that the small target of interest exists completely in the image slice). The four vertex images of SAS image I overlap on two sides, the remaining edge images overlap on three sides, and the remaining image slices overlap on four sides. For example, the slice file name in the upper left corner is 3200-500-500-450.jpg, which means r is 3200, c is 500, h is 500, and s is 450.

[0114] Image Slice I r,c,h,s SSD_MV3P generates the detection results of small targets of interest, including target type, probability and location information, which are recorded as:

[0115]

[0116] Among them B r,c,i,j Represents the detection information of the image slice, r represents the column coordinate in SAS image I, c represents the row coordinate in SAS image I, and i represents the type of the target.

[0117] Test result B r,c,i The overlapping suggestion boxes are removed by the non-maximum suppression algorithm Soft-NMS, and finally the bounding box with the highest score after regression correction at each position in each category is obtained, which is recorded as:

[0118]

[0119] Where S is the candidate box with the highest score in the current cycle, b r,c,i is the detection information at the current position, iou represents the intersection-over-union function, T t is the IoU threshold, It is the detection result of the small target of interest after secondary non-maximum suppression, and is finally displayed in the SAS image.

[0120] Step 4: Display the target detection results and display the targets of interest.

[0121] Example 2

[0122] Embodiment 2 of the present invention proposes a synthetic aperture sonar image underwater small target detection system, which is implemented based on the method of embodiment 1. The system includes:

[0123] 1. A processing module, which is used to process the acoustic echo data received by the array element to obtain a synthetic aperture sonar image, and obtain a number of image slices that meet the size requirements through a redundant cutting algorithm; it includes a synthetic aperture sonar submodule and an image redundant cutting submodule; wherein,

[0124] The synthetic aperture sonar submodule is used to process the received array data to obtain a real-time synthetic aperture sonar image;

[0125] The image cutting submodule is used to perform redundant cutting on synthetic aperture sonar images to ensure the integrity of small underwater targets;

[0126] 2. Platform deployment module, used to deploy the pre-built and trained small object detection model to the edge computing platform;

[0127] 3. A detection output module is used to input image slices one by one into a pre-established and trained small target detection model to obtain the target detection result of each slice, and then use secondary non-maximum suppression to filter out repeated recognition results of the same target in the slice, and finally realize underwater small target detection in synthetic aperture sonar images; it includes: a secondary non-maximum suppression submodule, which is used to filter out redundant recognition results of targets at the same position; and a result output submodule, which is used to display and output underwater small targets.

[0128] The small target detection model adopts an improved SSD network to achieve lightweight target detection. The improved SSD network includes a basic feature extraction network, and the basic feature extraction network adopts an improved feature extraction module.

[0129] The training process also includes: data set creation module and model training module.

[0130] 4. Dataset creation module

[0131] It includes: data collection submodule, data annotation submodule and target detection data set production submodule; among which,

[0132] The data acquisition submodule acquires underwater synthetic aperture sonar images from real environments;

[0133] The data annotation submodule annotates small underwater targets in synthetic aperture sonar images according to task requirements;

[0134] The target detection dataset production submodule randomly divides the data into training set and test set according to the standard target detection dataset format.

[0135] 5. Model training module

[0136] Includes: parameter setting submodule and model testing submodule.

[0137] The parameter setting submodule is used to complete the parameter initialization work required for model training;

[0138] The model testing module submodule is used to monitor the model training status in real time.

[0139] The following is a further description of the technical effects of the present invention in conjunction with simulation experiments:

[0140] The experimental platform operating system is RedHat 7.5, the environment management software is Anaconda3, the deep learning framework is Torch1.3.1 and Torchvision 0.4.2, etc.; BatchSize is equal to 64, the learning rate is equal to 0.001, and all models are trained and tested on GPU. In order to test the detection and recognition performance of HRSSD for small targets of interest in high-resolution and large-size SAS images, a SAS small target of interest dataset, SST (Sonar Small Targets), was established. SST contains three types of targets: non-targets, suspected objects, and cylindrical objects, with a total of 2490 images, collected from Qiandao Lake, Danjiangkou and other places.

[0141] In the HRSSD framework, this experiment compares and analyzes the performance differences of SE, ECA, CBMA and the BSC attention component of this paper on the SST dataset. The feature extraction module of SSD-MV3 is IRB + , using SE attention component; the feature extraction module of SSD-MV3E is IRB_E, using ECA attention component; the feature extraction module of SSD-MV3C is IRB_C, using CBAM attention component; the feature extraction module of SSD-MV3P is IRB ++, using the BSC attention component. The mAP value, parameter size, and average detection time of the detection model for the small target dataset SST of interest are recorded when the detection model is iterated 1000 times.

[0142] Table 1 Effect of different attention mechanisms on SSD-MV3P performance

[0143]

[0144] From Table 1, we can find that the detection accuracy of SSD-MV3P is 4.39%, 2.35% and 2.51% higher than that of SSD-MV3, SSD-MV3E and SSD-MV3C, respectively. The detection accuracy of SSD-MV3E and SSD-MV3C is close. The model parameters are 4.6MB, 7.1MB and 5.7MB higher than those of SSD-MV3, SSD-MV3E and SSD-MV3C, respectively. The detection time is 23.81ms, 24.94ms and 13.29ms higher than those of SSD-MV3, SSD-MV3E and SSD-MV3C, respectively. Considering that the target model and operation time are running at a relatively low level, the real-time requirements of small target detection of interest in SAS images can be met. SSD-MV3P is more suitable for small target detection and recognition of interest in SAS images.

[0145] In order to verify the effectiveness of HRSSD, HRSSD is used to detect and identify small targets of interest in high-resolution and large-size SAS images in three scenarios. The detection results are shown in the following table. Figure 5 , Figure 6 and Figure 7 As shown in the figure, the HRSSD algorithm based on SSD-MV3P can accurately detect and identify small targets of interest, and the detection results are complete and non-repetitive.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A method for detecting underwater small targets using synthetic aperture sonar images, the method comprising: Step 1) processing the acoustic echo data received by the array element to obtain a synthetic aperture sonar image, and obtaining a number of image slices that meet the size requirements through a redundant cutting algorithm; Step 2) input the image slices one by one into the pre-established and trained small target detection model to obtain the target detection result of each slice, and then use the secondary non-maximum suppression to screen the repeated recognition results of the same target in the slice, and finally realize the underwater small target detection of synthetic aperture sonar image; The small target detection model uses an improved SSD network to achieve lightweight target detection, and the improved SSD network includes a basic feature extraction network, and the basic feature extraction network uses an improved feature extraction module; The step 1) specifically includes: The acoustic echo data received by the array element is processed to obtain a synthetic aperture sonar image I with a size of N×N; After being processed by the redundant cutting algorithm, several image slices I are obtained. r,c,h,s , where r represents the column coordinate of the image slice in I, c represents the row coordinate of the image slice in I, h represents the size of the image slice, s represents the cutting step of the image slice, adjacent image slices overlap in the horizontal and vertical directions, and the single-side overlap ratio is a set value to ensure that the small target of interest exists completely in the image slice. The four vertex images of I overlap on two sides, the edge image overlaps on three sides, and the remaining image slices overlap on four sides; The small target detection model includes a basic feature extraction network, an additional feature extraction network, a candidate box generation module and a detection result prediction module connected in sequence, wherein: The basic feature extraction network includes two improved feature extraction modules, which output feature layers of two sizes respectively; The additional feature extraction network includes 4 convolutional layers, which respectively output feature layers of 4 sizes; The candidate box generation module is used to extract candidate boxes of different numbers and sizes from the feature layers of 6 sizes according to predefined scales and aspect ratios; The detection result prediction module is used to determine the type and position of the target in the candidate frame and optimize the target in the candidate frame using a non-maximum suppression algorithm; The processing process of the detection result prediction module specifically includes: For the detection information of the image slice B r,c,i,j , the overlapping suggestion boxes are removed by the secondary non-maximum suppression algorithm Soft-NMS, and finally the bounding box with the highest score after regression correction at each position in each category is obtained, which is the small target detection result Where r represents the column coordinate of the image slice in I, c represents the row coordinate of the image slice in I, and i represents the type of the target; where S i is the candidate box with the highest score in the current cycle, b r,c,i is the detection information at the current position, iou represents the intersection-over-union function, T t is the IoU threshold; The input of the improved feature extraction module is feature D, and the output is for: Wherein, D' is the output of the first branch, the first branch is a sequentially connected expansion layer, a segmentation layer, three parallel branches, a splicing layer, a channel selection component and a compression layer, the expansion layer and the compression layer both include a BN layer and a ReLU layer, the parallel branch includes a sequentially connected multi-scale component and a space selection component, wherein, The expansion layer is responsible for expanding the input feature channel; The segmentation layer is responsible for channel cutting; The multi-scale component is responsible for multi-scale decomposition of input information; The spatial selection component is responsible for selecting useful spatial information in the multi-scale layer; The splicing layer is responsible for channel integration; The channel selection component is used to select channels containing important information by learning weights; The compression layer is responsible for compressing the feature channel into a number consistent with the input feature; The processing process of the feature extraction module specifically includes: The input feature D passes through the expansion layer, and its output feature D ex for: D ex =F ex (D),D∈Φ H×H×M Among them, F ex This is a channel expansion operation. The convolution kernel size of the expansion layer is 1×1, and the number of output channels is the number of channels after expansion M. + ; D ex Enter the segmentation layer and the multi-scale components of each branch. The multi-scale component of the kth branch outputs D k branchs for: in, The features after segmentation are evenly segmented into three parallel branches; The output feature map size is H×H and the number of channels is M + / 3,F branchs It is a multi-scale depth-separable dilated convolution operation. D k branchs Input space selection component, whose output features Satisfy the following formula: in, The output feature map size is H×H and the number of channels is M + / 3; is the spatial weight of the kth branch, σ is the activation function, f 3*3 is a convolution operation with a convolution kernel size of 3×3, F concat is the channel splicing operation, AvgPool is the average pooling operation, and MaxPool is the maximum pooling operation; Use the splicing layer to 3 Concatenate in the channel dimension, the concatenated output feature D concat for: Among them, D concat The size of the output feature map is H×H, and the number of channels is M + , F concat It is a channel splicing operation; D concat Input channel selection component, whose output feature D chanel for: s chanel =MLP(AvgPool(D concat ))+MLP(MaxPool(D concat )) D chanel =s chanel ·D concat Among them, D chanel The size of the output feature map is H×H, and the number of channels is M + ,s chanel D concat The channel weight coefficient, MLP is a shared fully connected network; Right D chanel Channel compression is performed, and the characteristic D' after channel compression is: D'=F sq (D chanel ),D'∈Φ H×H×M Among them, the size of the output feature map of D' is H×H, and the number of channels is equal to the number of channels of the input feature D, which is M, F sq For channel compression operation, the convolution kernel size is 1×1.

2. The method for detecting underwater small targets using synthetic aperture sonar images according to claim 1, characterized in that: The method also includes a small target detection model training step, specifically including: Collect raw sonar data from real underwater environments, process them to obtain synthetic aperture sonar images of standard size, and annotate them to generate training sets; The training set data is input into the improved SSD network in sequence until the training requirements are met, and a trained small target detection model is obtained.

3. A system based on the method for detecting underwater small targets using synthetic aperture sonar images according to claim 1, characterized in that: The system comprises: A processing module is used to process the acoustic echo data received by the array element to obtain a synthetic aperture sonar image, and obtain a number of image slices that meet the size requirements through a redundant cutting algorithm; A platform deployment module for deploying pre-built and trained small object detection models to embedded platforms; and The detection output module is used to input the image slices one by one into the pre-established and trained small target detection model to obtain the target detection results of each slice, and then use the secondary non-maximum suppression to screen the repeated recognition results of the same target in the slice, and finally realize the underwater small target detection of synthetic aperture sonar images; The small target detection model adopts an improved SSD network to achieve lightweight target detection. The improved SSD network includes a basic feature extraction network, and the basic feature extraction network adopts an improved feature extraction module.

Images