Underwater target detection method and system based on multi-beam foresight sonar image

Abstract

The invention provides an underwater target detection method and system based on a multi-beam foresight sonar image, and belongs to the technical field of underwater image processing. The method comprises the following steps: acquiring a target sonar image by adopting a multi-beam foresight sonar device, extracting a multi-scale feature map by adopting an improved detection model based on a convolutional neural network, predicting and classifying bounding boxes, and outputting position coordinates and category probability result information containing each detection target. According to the invention, a dynamic perception patch attention mechanism is introduced into the improved convolutional neural network, and a multi-branch feature extraction strategy and a special small target detection head are combined, so that the detection capability of the small target is enhanced, and the purpose of improving the detection precision is achieved. In addition, sonar imaging and processing, data communication, display control and the like are integrated in an embedded platform, and portable real-time detection, data exchange and user control are achieved. According to the method, the detection precision of the small target in the underwater forward-looking multi-beam sonar image can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater sonar image processing, and in particular to an underwater target detection method and system based on multi-beam forward-looking sonar images. Background Art

[0002] With the rapid development of the global economy and science and technology, the development and utilization of marine resources have become increasingly important. The ocean not only contains rich natural resources, but also plays a key role in military strategy, scientific research and environmental protection. In order to better explore and utilize the ocean, countries have increased their investment in research on ocean detection technology. Among the many ocean detection technologies, Multibeam Forward-Looking Sonar (MFLS) can provide detailed seabed topography and underwater target information due to its high resolution and real-time imaging capabilities, and is widely used in underwater navigation, target detection, seabed mapping and other fields.

[0003] In recent years, deep learning technology has made breakthrough progress in the field of computer vision, especially in target detection. Through an end-to-end learning mechanism, deep learning-based methods can automatically extract image features and classify them, overcoming the limitations of hand-designed feature extractors and classifiers in traditional methods. However, deep learning methods also face challenges when applied to multi-beam forward-looking sonar images. On the one hand, training deep learning models requires a large amount of labeled data, while high-quality underwater sonar image datasets are relatively scarce; on the other hand, sonar images have unique characteristics, such as low signal-to-noise ratio, low resolution, and target sparsity. Directly applying existing deep learning models often makes it difficult to achieve ideal detection results. Therefore, how to combine the characteristics of multi-beam forward-looking sonar to design an efficient and accurate target detection method has become a problem that needs to be solved urgently. Summary of the invention

[0004] In order to solve the technical problems existing in the above-mentioned background technology, the purpose of the present invention is to provide an underwater target detection method and system based on multi-beam forward-looking sonar images. Based on convolutional neural networks, the detection network is improved by introducing an attention mechanism, and the detection head of the backbone network is optimized. In particular, a new detection head is introduced for small target detection to improve detection accuracy and efficiency. At the same time, by streamlining and optimizing the model structure and parameter quantity, the system is embedded in a microprocessor device to realize real-time detection and improve portability and practicality.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] In one aspect, the present invention provides an underwater target detection system based on multi-beam forward-looking sonar images, comprising a power module,

[0007] Multi-beam forward-looking sonar device, used to image the target waters and obtain original seabed sonar images;

[0008] A microprocessor is used to run the improved detection model based on convolutional neural network, obtain target detection results and control other modules, wherein the target detection results include the position, size and classification information of the target; the improved detection model based on convolutional neural network is to introduce a dynamic perception patch attention mechanism into the backbone network of the convolutional neural network to replace the convolution operation in the original encoder and decoder; capture multi-scale features through a multi-branch feature extraction strategy, and enhance the extracted multi-scale features through dynamic feature fusion and attention mechanism; adopt a neck network structure based on feature pyramid network and depth separable convolution;

[0009] Data storage module, used to store original seabed sonar images, preprocessed data, trained network parameters and intermediate result information;

[0010] A communication module, used to transmit the target detection result to a remote monitoring center and receive instructions from a remote location;

[0011] Display and control module, used to display test results and send control instructions;

[0012] The multi-beam forward-looking sonar device, the data storage module, the communication module, the display and control module and the power supply module are all connected to the microprocessor.

[0013] Preferably, the display and control module is specifically used to:

[0014] The user instructions are received in real time and control signals are sent to the microprocessor through a communication interface; the user instructions include: adjusting the posture of the multi-beam forward-looking sonar device and selecting a specific target for continuous tracking.

[0015] Preferably, a towed wired communication method is adopted between the microprocessor and the multi-beam forward-looking sonar device.

[0016] Preferably, the improved detection model based on convolutional neural network is obtained by:

[0017] Build an improved convolutional neural network; the improved convolutional neural network is to introduce a dynamic perception patch attention mechanism in the backbone network of the convolutional neural network to replace the convolution operation in the original encoder and decoder; capture multi-scale features through a multi-branch feature extraction strategy, and enhance the extracted multi-scale features through dynamic feature fusion and attention mechanism; adopt a neck network structure based on feature pyramid network and depth separable convolution;

[0018] Collect multi-beam forward-looking sonar images and annotate target locations and categories to form a training data set;

[0019] The training data set is input into the improved convolutional neural network for training, and cross-validation is used to avoid overfitting. At the same time, through performance feedback on the validation set, the network hyperparameters are continuously optimized, and the network hyperparameters with the best performance are retained to obtain an improved detection model based on the convolutional neural network.

[0020] Preferably, the dynamic-aware patch attention mechanism captures multi-scale features of the input image through a multi-branch feature extraction strategy, including:

[0021] Dynamic local branch, used to extract local features of input images;

[0022] Dynamic global branch, used to extract global features of input images;

[0023] Dynamic serial convolution branch, used to extract features of input image sequence;

[0024] The dynamic local branch data processing process is as follows:

[0025] The input preprocessed sonar image Adjust the number of channels through 1×1 convolution to generate a new feature tensor ;

[0026] Adopting adaptive patch size The feature tensor Split into a set of spatially contiguous patches ;

[0027] Channel-wise averaging is performed on each patch to obtain , and use the feedforward neural network for linear calculation to get ;

[0028] The activation function is used to obtain the probability distribution of the spatial dimension of the linear calculation feature, and feature selection is performed on each patch based on the probability distribution to obtain the feature ;

[0029] Features Perform reshape and interpolation operations to generate local features ;

[0030] The dynamic global branch data processing process is as follows:

[0031] For the feature tensor Perform global average pooling to obtain vector ;

[0032] By linear transformation of vector Reduce the dimension, then introduce nonlinearity through the activation function Sigmoid, and finally increase the dimension back to the original number of channels to obtain the vector ;

[0033] Pair Vector Apply the activation function Softmax for normalization to obtain the weights of each channel ;

[0034] Multiply the weight of each channel by The feature map of the corresponding channel is then added element by element to obtain the global feature after adaptive aggregation ;

[0035] For global features Perform dynamic attention calculation to obtain the dynamic attention matrix ;

[0036] The dynamic attention matrix and the feature tensor Multiply them together to get the final global feature ;

[0037] The dynamic serial convolution branch data processing process is:

[0038] According to the feature tensor Perform global average pooling and get ;

[0039] Will Input the multi-layer perceptron MLP and get a selection vector indicating the size and number of convolution kernels ;

[0040] Randomly initialize and generate three convolution kernel templates , , , based on the selection vector , by combining convolution kernel templates of different sizes in a weighted summation manner, we get the convolution kernel , , ;

[0041] Three output features are generated through a 3×3 convolutional layer , and , expressed as: , , ,

[0042] The three output features are integrated through an adaptive fusion mechanism , and Perform dynamic fusion to generate serial convolution features ;

[0043] The local features , global features and serial convolution features Add together to get the output features .

[0044] Preferably, the dynamic perception patch attention mechanism enhances the extracted multi-scale features through dynamic feature fusion and attention mechanism, and the data processing process is:

[0045] The output features are processed through a one-dimensional dynamic channel attention mechanism Process and generate channel attention features ;

[0046] Channel attention features through two-dimensional dynamic spatial attention mechanism Processing to generate spatial attention features ;

[0047] The spatial attention feature After ReLU activation function and batch normalization, the final output feature .

[0048] Preferably, the data processing process of the neck network structure based on the feature pyramid network and the depth-separable convolution is:

[0049] Extract multi-scale feature maps from the backbone network: , , , , corresponding to feature maps at different levels;

[0050] The feature maps of different levels are fused through the feature pyramid network, which is expressed as:

[0051] ;

[0052] in, Represents an upsampling operation;

[0053] Will Perform depth-wise separable convolution to get the output ; The depth-wise separable convolution includes depth-wise convolution and point-wise convolution;

[0054] The squeeze excitation module introduced in SENet is Process and get output .

[0055] Preferably, the improved detection head operates as follows:

[0056] ,

[0057] in, and Represent classification and regression convolutional layers respectively.

[0058] In a second aspect, the present invention provides an underwater target detection method based on a multi-beam forward-looking sonar image, which is implemented by using the above-mentioned underwater target detection system based on a multi-beam forward-looking sonar image, and the method comprises:

[0059] Use a multi-beam forward-looking sonar device to obtain sonar images of the target waters;

[0060] Preprocessing the acquired sonar images;

[0061] The preprocessed sonar image is input into the improved convolutional neural network-based detection model, multi-scale feature map extraction is performed on the input sonar image, and bounding box prediction and classification are performed, and the output includes the location coordinates and category probability result information of each detected target;

[0062] Filter the valid target frame, obtain the valid target detection result, and transmit it to the display and control module for display.

[0063] Preferably, the preprocessing of the acquired sonar image comprises:

[0064] Performing denoising processing on the sonar image of the target water area based on wavelet transform;

[0065] Normalize the denoised image;

[0066] The screening effective target frame includes:

[0067] Valid targets are screened out according to the confidence threshold in the output results, and the non-maximum suppression algorithm is used to remove redundant boxes.

[0068] Beneficial effects of the present invention:

[0069] The present invention provides a method and system for underwater target detection based on multi-beam forward-looking sonar images, which significantly improves the accuracy of small target recognition by combining convolutional neural networks with dynamic perception patch attention mechanisms. At the same time, a dedicated small-size target detection head is designed to capture details and maintain high computing efficiency, and the model structure is optimized to make it suitable for embedding in microprocessor devices to achieve portable real-time detection. The system supports human-computer interaction through a communication interface, enhances the convenience and flexibility of operation, and provides an efficient solution for underwater detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1A schematic diagram of the architecture of an underwater target detection system based on multi-beam forward-looking sonar images provided by an embodiment of the present invention;

[0071] Figure 2 A schematic flow chart of an underwater target detection method based on multi-beam forward-looking sonar images provided by an embodiment of the present invention;

[0072] Figure 3 A schematic diagram of the network structure of the dynamic perception patch attention mechanism provided in an embodiment of the present invention;

[0073] Figure 4 A schematic diagram of a neck network structure based on a multi-scale receptive field provided in an embodiment of the present invention;

[0074] Figure 5 This is an example diagram of the detection results provided in the embodiments of the present invention. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0076] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0077] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0078] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0079] It should be emphasized here that the step marks mentioned below are not intended to limit the order of the steps, but it should be understood that the steps can be executed in the order mentioned in the embodiment, or in a different order from the embodiment, or several steps can be executed simultaneously.

[0080] The present invention provides an underwater target detection system based on multi-beam forward-looking sonar (MFLS) images. The system optimizes a convolutional neural network (CNN) model and introduces a specific mechanism to improve the detection accuracy of small targets while ensuring the portability and real-time performance of the system.

[0081] Based on this, the underwater target detection system based on multi-beam forward-looking sonar (MFLS) images provided by the present invention is shown in FIG. Figure 1 , including the following required hardware equipment:

[0082] Multi-beam forward-looking sonar data acquisition: Use a multi-beam forward-looking sonar device to image the target waters and obtain high-resolution seabed sonar images.

[0083] Microprocessor: The ARM-based microprocessor is used to run the improved convolutional neural network (CNN)-based detection model to obtain target detection results and control other components. The target detection results include the location, size, and classification information of the target.

[0084] Data storage module: used to store information such as raw seabed sonar data, preprocessed data, trained model parameters, and intermediate results.

[0085] Power module: Provides stable power supply for the entire system to ensure long-term operation.

[0086] Communication module: transmits detected data to the remote monitoring center through wired connection and receives instructions from the remote.

[0087] Display and control module: used to display the test results and provide a user interface for interaction, allowing operators to view detailed information and send control instructions.

[0088] In the present invention, the multi-beam forward-looking sonar device includes multiple transmitting and receiving units, which can generate high-resolution seabed sonar images. In addition, in view of the underwater noise and sound wave propagation characteristics, the original seabed sonar image data needs to be pre-processed and denoised to improve the data quality. In addition, the image standardization process ensures the consistency of the model input data.

[0089] Furthermore, the embedded computing platform also includes a dedicated hardware accelerator for real-time processing and analysis of sonar images.

[0090] Furthermore, the display and control module can receive user instructions in real time and send control signals to the underwater microprocessor through the communication interface to adjust the posture of the multi-beam forward-looking sonar device equipped with the microprocessor or perform specific functions.

[0091] In the present invention, the improved convolutional neural network (CNN) structure is as follows:

[0092] 1. Introducing Dynamic-Aware Patch Attention Mechanism

[0093] A dynamic-aware patch attention mechanism is integrated into the CNN backbone network, enabling the model to automatically optimize its internal weights to enhance sensitivity to small objects and overall recognition accuracy.

[0094] In the sonar target detection task, since multiple downsampling operations may lead to the loss of key information of small targets, the dynamic patch attention mechanism is designed to replace the convolution operations in traditional encoders and decoders, and a multi-branch feature extraction strategy is adopted, which can capture target features from different scales and levels, thereby improving the detection accuracy of small targets.

[0095] like Figure 3 As shown, in the present invention, a dynamic-aware patch attention mechanism (DPA) is adopted to capture multi-scale features through a multi-branch feature extraction strategy, and multi-scale features are fused through dynamic feature fusion and attention mechanism to enhance the sensitivity to key information.

[0096] Specifically, the multi-branch feature extraction includes: dynamic local branch, dynamic global branch and dynamic serial convolution branch. The data processing process of each branch is as follows:

[0097] (1) Dynamic local branching

[0098] Input conditioning: Preprocessed sonar image given input , first adjust the number of channels through 1×1 convolution to generate a new feature tensor , expressed as:

[0099] (4)

[0100] Adaptive Patch Segmentation: Using Adaptive Patch Size Will Split into a set of spatially contiguous patches , expressed as:

[0101] (5)

[0102] Channel averaging and linear calculation: Channel averaging is performed on each patch to obtain , expressed as:

[0103] (6)

[0104] Linear calculation is performed using a feedforward neural network (FFN), expressed as:

[0105] (7)

[0106] Dynamic feature selection: An activation function is used to obtain the probability distribution of the spatial dimension of the linearly calculated features, and the weights are adjusted accordingly. The weights are expressed as:

[0107] (8)

[0108] Feature selection is performed on each patch and the output is: ,in is the sigmoid function,

[0109] (9)

[0110] Finally, local features are generated through reshape and interpolation operations: , expressed as:

[0111] (10)

[0112] (2) Dynamic global branch

[0113] Adaptive global feature aggregation: By introducing an adaptive aggregation mechanism, the aggregation method is dynamically adjusted according to the importance of global features to generate global features.

[0114] (11)

[0115] The operation of the adaptive aggregation mechanism (AdaptiveAgg) is as follows:

[0116] ① Apply global average pooling (GAP) to F′ to compress the spatial information of each channel into a scalar value. The result is a vector G of shape (C′) representing the global features of each channel.

[0117] in, ;

[0118] ② Calculate channel attention and assign weights to each channel, first by reducing the dimension, then introducing nonlinearity through the activation function Sigmoid, and finally increasing the dimension back to the original number of channels. The specific operations are as follows:

[0119] The global features are transformed into Mapping to a lower dimensional space , where r is the middle dimension and its value is C′ / 16. This step compresses the feature space so that the model can better focus on important features.

[0120] ,

[0121] in, is the weight matrix of the first layer, is the bias vector of the first layer, It is the feature representation after dimensionality reduction.

[0122] The activation function ReLU is applied to introduce nonlinearity so that the model can handle nonlinear relationships.

[0123] ,

[0124] in, It is the feature representation after ReLU activation.

[0125] Again, the features are transformed from low-dimensional Mapping back to high latitude .

[0126] ,

[0127] in, is the weight matrix of the second layer, is the bias vector of the second layer, It is the feature representation after dimensionality increase.

[0128] Apply the activation function Softmax to normalize the weights so that they are in the interval [0,1] and the sum of all weights is 1

[0129] ,

[0130] in, ,and ,

[0131] This ensures that the output weights are normalized probability distributions so that they can be used as the importance weights of each channel.

[0132] ③ Weighted summation: multiply the weight of each channel by the corresponding feature map, and then add them element by element. This can highlight important feature maps while suppressing unimportant feature maps.

[0133] ,

[0134] in, is the feature map of the cth channel, yes The feature map of the cth channel;

[0135] The final global features It is the global feature after adaptive aggregation.

[0136] Dynamic attention calculation: Calculate the dynamic attention matrix , a dynamic weight adjustment mechanism is introduced, which is expressed as follows:

[0137] (12)

[0138] The dynamic weight adjustment mechanism here (DynamicAttention) is to build a dynamic attention matrix According to the above calculation, , construct a Tensors of the same size , where the weight of each position comes from the corresponding , that is, ;

[0139] Finally, the global features are generated: , expressed as:

[0140] (13)

[0141] (3) Dynamic serial convolution branch

[0142] Adaptive convolution kernel: Dynamically select the size and number of convolution kernels according to the characteristics of the input feature map to generate an adaptive convolution kernel , expressed as:

[0143] (14)

[0144] The specific operations are as follows:

[0145] ① Perform global average pooling (GAP) on the feature map F′ to capture important information in the spatial and channel dimensions, and calculate the average value on each channel to reflect the performance of each channel on the entire image:

[0146] ,

[0147] in and are the height and width of the feature map, respectively. A vector representing all channels at position (i, j).

[0148] ② Determine the size and number of convolution kernels, use a multi-layer perceptron MLP, receive the above analysis results as input, and output a selection vector S indicating the size and number of convolution kernels:

[0149] ,

[0150] vector Each element of corresponds to the weight of a specific convolution kernel configuration.

[0151] ③ Dynamically select convolution kernel parameters and use random initialization to generate three convolution kernel templates , , , based on the selection vector , by combining convolution kernel templates of different sizes in a weighted summation manner to generate the final convolution kernel , , , which is expressed as follows:

[0152] ,

[0153] is from the selection vector The weight of is an additional weight term used to adjust the influence of convolution kernel templates of other sizes.

[0154] Convolution operation: Generate three different outputs through a 3×3 convolution layer , , , expressed as:

[0155] (15)

[0156] (16)

[0157] (17)

[0158] The three output features are dynamically fused through the adaptive fusion mechanism to finally generate serial convolution features: , expressed as:

[0159] (18)

[0160] Finally, the local features, global features and serial convolution features are added together to obtain:

[0161] .

[0162] In the present invention, the dynamic feature fusion and attention mechanism data processing process is as follows:

[0163] Through a one-dimensional dynamic channel attention map right Process and generate channel attention features , expressed as:

[0164] (19)

[0165] (20)

[0166] By capturing the dependencies between different channels, the model can better focus on more important feature channels. The specific operations are as follows:

[0167] ① Global pooling, respectively, the input feature map Perform global average pooling (GAP) and global maximum pooling (GMP) to obtain statistics on each channel.

[0168] ,

[0169] ,

[0170] ② Feed the results of GAP and GMP into a shared multi-layer perceptron (MLP), which contains two fully connected layers for compressing and restoring the number of channels.

[0171] in ,

[0172] σ is the activation function ReLU, and are the weight matrices of the first and second layers respectively, It is another activation function sigmoid.

[0173] ③Concatenate the outputs of GAP and GMP, and then generate the channel attention map through the sigmoid function ,This graph is used to adjust the importance of each channel of the original feature map.

[0174] in ,

[0175] ④ Channel attention map Applied to the original feature map , get the channel attention feature ,

[0176] .

[0177] Through the two-dimensional dynamic spatial attention map Channel attention features Processing to generate spatial attention features , expressed as:

[0178] (twenty one)

[0179] (twenty two)

[0180] Focusing on information in the spatial dimension helps the model focus on key areas in the image. The specific operations are as follows:

[0181] ① Feature map conversion, first input feature map The channel dimension is compressed to a single channel by average pooling, where ,

[0182] ② Use a 7×7 convolution kernel to perform a convolution operation on the compressed feature map to capture a wider range of spatial context information, where ,

[0183] ③Generate a spatial attention map and generate an attention map through the sigmoid function , which is used to adjust the original feature map The importance of each position in ,

[0184] ④The spatial attention map , applied to channel attention features , and obtain the final spatial attention feature .

[0185] After ReLU activation function and batch normalization, the final output is ,

[0186] (twenty three)

[0187] In the task of detecting small targets in forward-looking multi-beam sonar images, the dynamic perception patch attention mechanism (DPA) is suitable for dealing with the problem of low resolution and high noise in sonar images. DPA uses a multi-branch feature extraction strategy, dynamic local branches and dynamic global branches to effectively capture multi-scale features and enhance the detection ability of small targets. The application of dynamic feature fusion and attention mechanism further improves the model's sensitivity to key information and makes small target detection more accurate. The specific structure diagram is shown in the figure below. Figure 3 shown.

[0188] 2. Design of a neck network structure based on multi-scale receptive fields

[0189] In the present invention, a detection head module for small targets is designed in the CNN neck network structure, and small targets are accurately captured through multi-scale receptive fields. At the same time, the model structure is simplified and the number of parameters is reduced, the calculation efficiency of the model is improved, and it is more suitable for running on embedded platforms.

[0190] The neck network structure based on multi-scale receptive field is as follows Figure 4 As shown, the data processing process is as follows:

[0191] (1) Feature extraction: Multi-scale feature maps are extracted from the backbone network (ResNet) as follows: , , , , corresponding to feature maps at different levels;

[0192] (2) Feature Pyramid Network (FPN): FPN fuses feature maps of different levels through top-down paths and lateral connections, expressed as:

[0193] (twenty four)

[0194] in, Represents an upsampling operation using bilinear interpolation;

[0195] The highest level feature map P 4 Directly from the last layer feature map C of the backbone network 4 Get;

[0196] Combining the above processing methods, the complete FPN feature fusion formula is:

[0197] ;

[0198] (3) Depthwise separable convolution: Depthwise separable convolution consists of depthwise convolution and pointwise convolution, which reduces the number of parameters and calculations. The input feature map , the specific operation of depth-wise separable convolution is as follows:

[0199] ,

[0200] in, Denotes a deep convolution, which performs convolution on each channel of the input separately, rather than combining all channels together as in standard convolution. This generates a new set of feature maps , which is the same as the number of channels, but the number of parameters is greatly reduced.

[0201] ,

[0202] in, Represents point-wise convolution. When processing a large number of feature maps generated by multi-beam forward-looking sonar, point-wise convolution is used to adjust the number of channels to ensure that the output feature map can meet the needs of subsequent tasks.

[0203] (4) Attention mechanism: The squeeze excitation module in SENet is introduced to enhance the attention to small targets. The squeeze excitation module can dynamically adjust the importance of each channel, helping the model to better focus on key areas such as small targets, thereby improving detection accuracy. The operation of the squeeze excitation module of the input feature map Y is as follows:

[0204] (26)

[0205] in, represents global average pooling, and They represent the fully connected layer, represents the activation function, Represents element-wise multiplication.

[0206] The specific role of the squeeze incentive module:

[0207] Squeeze: First, the spatial information of each channel is compressed into a scalar value through global average pooling (GAP), thereby reducing the computational complexity and preserving the overall feature representation of each channel.

[0208] Excitation: Use two fully connected layers (FC1 and FC2) with a ReLU activation function in between to learn the importance weights of each channel, and finally normalize these weights to the [0, 1] interval through the Sigmoid activation function σ.

[0209] Adjustment: These weights are multiplied element by element Apply it back to the feature map Y to enhance important features and suppress irrelevant features.

[0210] 3. Design a special detection head, use a smaller anchor box to match small targets, and input feature maps The detection head operates as follows:

[0211] (27)

[0212] in, and Represent classification and regression convolutional layers respectively.

[0213] Based on the improved convolutional neural network, a large number of representative forward-looking multi-beam sonar images are collected and annotated to form a training data set, which contains a variety of underwater scenes and targets.

[0214] The training data set is input into the above-mentioned improved convolutional neural network for training, and cross-validation is used to avoid overfitting. At the same time, through performance feedback on the validation set, the model hyperparameters are continuously optimized to obtain the optimal performance, and finally an improved detection model based on convolutional neural network is obtained.

[0215] It should be further explained that after completing the above-mentioned model training, the trained detection model is converted into a format suitable for microprocessor operation and loaded into the microprocessor.

[0216] A dedicated interface between the microprocessor and the sonar sensor is configured to ensure efficient transmission of data streams. Due to the particularity of the underwater environment, a towed wired communication method is adopted to connect the microprocessor and the sonar sensor through a dedicated cable to ensure the stability and low latency of data transmission.

[0217] The microprocessor transmits the detected target position coordinates and category information to the user interface in real time through the output interface.

[0218] The user clicks on the target icon on the interface, and the microprocessor receives the user request, extracts the size, shape and confidence of the target from the stored detailed information, and displays this information through the user interface.

[0219] The user sends control instructions to the underwater device through the interface, and the microprocessor receives these instructions and processes them, including:

[0220] When a specific target is selected for continuous tracking, the microprocessor generates corresponding control signals based on the position coordinates of the target and sends them to the sonar device through the drag-type wired communication interface to adjust its direction and focal length to keep the target always in the center of the field of view;

[0221] When the user manually adjusts the direction of the sonar device, the microprocessor generates corresponding control signals based on the user's instructions and sends them to the sonar device through the drag-type wired communication interface to adjust its angle to obtain better field of view coverage.

[0222] Based on the above underwater target detection system, the present invention provides an underwater target detection method based on multi-beam forward-looking sonar images, see Figure 2 , the method comprising:

[0223] Use a multi-beam forward-looking sonar device to obtain sonar images of the target waters;

[0224] Preprocessing the acquired sonar images;

[0225] The preprocessed sonar image is input into the improved convolutional neural network-based detection model, multi-scale feature map extraction is performed on the input sonar image, and bounding box prediction and classification are performed, and the output includes the location coordinates, category probability and other information of each detected target;

[0226] Filter out valid targets based on the confidence threshold in the output results and remove redundant boxes;

[0227] The target detection results are transmitted to the display and control module through the communication interface for display.

[0228] It should be noted that, in the present invention, the improved convolutional neural network is trained in advance with a large number of sonar image data sets marked with target positions and categories to obtain a detection model based on the convolutional neural network.

[0229] It should be noted that the improved convolutional neural network structure and the process of training the detection model are as described above and will not be repeated here.

[0230] It should be noted that, in the present invention, valid targets are screened out according to the confidence threshold in the output result, and a non-maximum suppression algorithm is used to remove redundant boxes to ensure that only one best matching bounding box is retained for each type of target.

[0231] It should be noted that, in the present invention, after receiving the data, the display and control module presents clear target detection results on the screen, including but not limited to the location, size and classification information of the target.

[0232] In one embodiment of the present invention, the acquired sonar image is preprocessed, including denoising and standardization, wherein:

[0233] De-noising of multi-beam forward-looking sonar images based on wavelet transform, including:

[0234] Given a size of For an image f(x,y), the two-dimensional discrete wavelet transform is defined as follows:

[0235] (1)

[0236] in, Indicates Tier Location Wavelet coefficients in (horizontal, vertical or diagonal) directions; is the corresponding wavelet basis function.

[0237] Wavelet decomposition of the image: Perform three-level wavelet decomposition on the original image, and an approximate image will be obtained after each level of decomposition. and three detail images , , (representing details in horizontal, vertical and diagonal directions respectively).

[0238] Thresholding: Estimate the standard deviation of noise in flat areas to obtain , determine the threshold based on the noise standard deviation ,in The expression is:

[0239] (2)

[0240] Use soft thresholding to process detail coefficients:

[0241] (3)

[0242] For each detail image , , Soft thresholding is applied to all coefficients of .

[0243] Reconstruct the image: Use the wavelet coefficients after threshold processing to perform inverse wavelet transform to reconstruct the denoised image.

[0244] Figure 5 This is a specific implementation example of the method of the present invention. The left picture is the original image, i.e., the forward-looking multi-beam sonar image, and the right picture is the image display obtained after detection by this method. It shows good results in identifying the three targets in the image, namely, square cage, cube and tyre, and successfully achieves the expected goal.

[0245] The specific embodiments described above are only used to elaborate the technical solutions and implementation details of the present invention, but should not be understood as limiting the scope of the present invention. Without departing from the basic principles and core ideas of the present invention, those skilled in the art can appropriately adjust, modify or replace the technical features in the above embodiments according to actual needs. All these changes or improvements made based on the spirit of the present invention should be deemed to fall within the protection scope of the present invention.

Claims

1. An underwater target detection system based on multi-beam forward-looking sonar images, characterized in that: Including power module, Multi-beam forward-looking sonar device, used to image the target waters and obtain original seabed sonar images; A microprocessor is used to run the improved detection model based on convolutional neural network, obtain target detection results and control other modules, wherein the target detection results include the position, size and classification information of the target; the improved detection model based on convolutional neural network is to introduce a dynamic perception patch attention mechanism into the backbone network of the convolutional neural network to replace the convolution operation in the original encoder and decoder; capture multi-scale features through a multi-branch feature extraction strategy, and enhance the extracted multi-scale features through dynamic feature fusion and attention mechanism; adopt a neck network structure based on feature pyramid network and depth separable convolution; Data storage module, used to store original seabed sonar images, preprocessed data, trained network parameters and intermediate result information; A communication module, used to transmit the target detection result to a remote monitoring center and receive instructions from a remote location; Display and control module, used to display test results and send control instructions; The multi-beam forward-looking sonar device, the data storage module, the communication module, the display and control module and the power supply module are all connected to the microprocessor.

2. The underwater target detection system based on multi-beam forward-looking sonar images according to claim 1, characterized in that: The display and control module is specifically used for: The user instructions are received in real time and control signals are sent to the microprocessor through a communication interface; the user instructions include: adjusting the posture of the multi-beam forward-looking sonar device and selecting a specific target for continuous tracking.

3. The underwater target detection system based on multi-beam forward-looking sonar images according to claim 1, characterized in that: A towed wired communication mode is adopted between the microprocessor and the multi-beam forward-looking sonar device.

4. The underwater target detection system based on multi-beam forward-looking sonar images according to claim 1, characterized in that: The improved detection model based on convolutional neural network is obtained in the following way: Build an improved convolutional neural network; the improved convolutional neural network is to introduce a dynamic perception patch attention mechanism in the backbone network of the convolutional neural network to replace the convolution operation in the original encoder and decoder; capture multi-scale features through a multi-branch feature extraction strategy, and enhance the extracted multi-scale features through dynamic feature fusion and attention mechanism; adopt a neck network structure based on feature pyramid network and depth separable convolution; Collect multi-beam forward-looking sonar images and annotate target locations and categories to form a training data set; The training data set is input into the improved convolutional neural network for training, and cross-validation is used to avoid overfitting. At the same time, through performance feedback on the validation set, the network hyperparameters are continuously optimized, and the network hyperparameters with the best performance are retained to obtain an improved detection model based on the convolutional neural network.

5. The underwater target detection system based on multi-beam forward-looking sonar images according to claim 4 is characterized in that: The dynamic-aware patch attention mechanism captures multi-scale features of the input image through a multi-branch feature extraction strategy, including: Dynamic local branch, used to extract local features of input images; Dynamic global branch, used to extract global features of input images; Dynamic serial convolution branch, used to extract features of input image sequence; The dynamic local branch data processing process is as follows: The input preprocessed sonar image Adjust the number of channels through 1×1 convolution to generate a new feature tensor ; Adopting adaptive patch size The feature tensor Split into a set of spatially contiguous patches ; Channel-wise averaging is performed on each patch to obtain , and use the feedforward neural network for linear calculation to get ; The activation function is used to obtain the probability distribution of the spatial dimension of the linear calculation feature, and feature selection is performed on each patch based on the probability distribution to obtain the feature ; Features Perform reshape and interpolation operations to generate local features ; The dynamic global branch data processing process is as follows: For the feature tensor Perform global average pooling to obtain vector ; By linear transformation of vector Reduce the dimension, then introduce nonlinearity through the activation function Sigmoid, and finally increase the dimension back to the original number of channels to obtain the vector ; Pair Vector Apply the activation function Softmax for normalization to obtain the weights of each channel ; Multiply the weight of each channel by The feature map of the corresponding channel is then added element by element to obtain the global feature after adaptive aggregation ; For global features Perform dynamic attention calculation to obtain the dynamic attention matrix ; The dynamic attention matrix and the feature tensor Multiply them together to get the final global feature ; The dynamic serial convolution branch data processing process is: According to the feature tensor Perform global average pooling and get ; Will Input the multi-layer perceptron MLP and get a selection vector indicating the size and number of convolution kernels ; Randomly initialize and generate three convolution kernel templates , , , based on the selection vector , by combining convolution kernel templates of different sizes in a weighted summation manner, we get the convolution kernel , , ; Three output features are generated through a 3×3 convolutional layer , and , expressed as: , , , The three output features are integrated through an adaptive fusion mechanism , and Perform dynamic fusion to generate serial convolution features ; The local features , global features and serial convolution features Add together to get the output features .

6. The underwater target detection system based on multi-beam forward-looking sonar images according to claim 5, characterized in that: The dynamic perception patch attention mechanism enhances the extracted multi-scale features through dynamic feature fusion and attention mechanism. The data processing process is: The output features are processed through a one-dimensional dynamic channel attention mechanism Process and generate channel attention features ; Channel attention features through two-dimensional dynamic spatial attention mechanism Processing to generate spatial attention features ; The spatial attention feature After ReLU activation function and batch normalization, the final output feature .

7. The underwater target detection system based on multi-beam forward-looking sonar images according to claim 6, characterized in that: The neck network structure based on feature pyramid network and depth separable convolution, the data processing process is: Extract multi-scale feature maps from the backbone network: , , , , corresponding to feature maps at different levels; The feature maps of different levels are fused through the feature pyramid network, which is expressed as: ； in, Represents an upsampling operation; Will Perform depth-wise separable convolution to get the output ; The depth-wise separable convolution includes depth-wise convolution and point-wise convolution; The squeeze excitation module introduced in SENet is Process and get output .

8. The underwater target detection system based on multi-beam forward-looking sonar images according to claim 7 is characterized in that: The improved detection head operates as follows: ， in, and Represent the classification and regression convolutional layers respectively.

9. A method for underwater target detection based on multi-beam forward-looking sonar images, characterized in that: The underwater target detection system based on multi-beam forward-looking sonar images according to any one of claims 1 to 8 is implemented, and the method comprises: Use a multi-beam forward-looking sonar device to obtain sonar images of the target waters; Preprocessing the acquired sonar images; The preprocessed sonar image is input into the improved convolutional neural network-based detection model, multi-scale feature map extraction is performed on the input sonar image, and bounding box prediction and classification are performed, and the output includes the location coordinates and category probability result information of each detected target; Filter the valid target frame, obtain the valid target detection result, and transmit it to the display and control module for display.

10. The underwater target detection method based on multi-beam forward-looking sonar images according to claim 9, characterized in that: The preprocessing of the acquired sonar image comprises: Performing denoising processing on the sonar image of the target water area based on wavelet transform; Normalize the denoised image; The screening effective target frame includes: Valid targets are screened out according to the confidence threshold in the output results, and the non-maximum suppression algorithm is used to remove redundant boxes.

Images