A lightweight target recognition method for underwater sonar images

Abstract

The application provides a lightweight target recognition method for underwater sonar images, and belongs to the technical fields of computer vision, deep learning and underwater intelligent perception. In view of problems such as low resolution, strong speckle noise, complex background echo, fuzzy target boundary and significant scale change of the sonar image, a lightweight recognition network with phased feature enhancement is constructed. The method introduces a cross-stage aggregation module CSA in the shallow feature extraction stage, improves the expression ability of target edge, texture and local structure information through channel shunting, multi-path feature aggregation and channel attention enhancement, introduces an improved inverted residual feature encoding module EMBConv in the deep semantic feature extraction stage, enhances the effective channel response and suppresses noise interference through an efficient channel attention module ECA, and introduces a hybrid spatial pyramid pooling module HSPP before classification output, fuses multi-scale context information through local multi-scale pooling, multi-dilation rate atrous convolution and global average pooling. The method can improve the accuracy and robustness of underwater sonar image target recognition, and is suitable for underwater robot autonomous detection, underwater target rapid recognition and marine environment intelligent perception and the like.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning, and underwater intelligent sensing, and particularly to a lightweight target recognition method, system, and computer-readable storage medium for underwater sonar images. This method can be applied to scenarios such as autonomous underwater robot exploration, intelligent sonar image analysis, rapid underwater target identification, marine resource exploration, underwater environmental monitoring, and marine engineering inspection. Background Technology

[0002] Sonar imaging technology is an important means of underwater sensing, capable of acquiring information about underwater targets and the environment in low visibility, turbid waters, and dark environments. Compared with optical imaging methods, sonar imaging is not dependent on lighting conditions and has strong environmental adaptability, thus it is widely used in tasks such as underwater robot detection, seabed topography mapping, underwater target identification, and marine engineering inspection.

[0003] However, the imaging mechanism of sonar images differs significantly from that of natural optical images. Affected by factors such as multipath reflection during sound wave propagation, environmental reverberation, medium absorption, equipment noise, and speckle interference, sonar images typically suffer from low resolution, blurred target boundaries, insufficient texture information, uneven brightness distribution, and strong background noise. These problems make it difficult for traditional image recognition methods to accurately extract effective target features, and also cause existing convolutional neural network models to exhibit insufficient recognition accuracy, weak generalization ability, and unstable noise resistance when directly transferred to sonar image recognition tasks.

[0004] Currently, lightweight convolutional neural networks have achieved good accuracy and efficiency in natural image classification tasks, reducing the number of model parameters and computational cost through methods such as depthwise separable convolution, inverse residual structures, and channel attention mechanisms. However, these networks are usually designed primarily for natural optical images, and when directly applied to underwater sonar image target recognition tasks, they still have the following shortcomings: First, the shallow feature extraction stage is insufficient in preserving weak edges, weak textures, and local structural information in sonar images; second, the deep feature extraction stage has limited ability to suppress noise channels and invalid background responses; third, there is a lack of context enhancement mechanisms for multi-scale sonar targets and complex background reverberation before classification output.

[0005] Therefore, it is necessary to design a lightweight target recognition method tailored to the characteristics of underwater sonar images, which can improve the recognition accuracy and robustness of low-resolution, high-noise, and weak-feature targets while ensuring the simplicity and ease of deployment of the model structure. Summary of the Invention

[0006] To address the shortcomings of existing sonar image target recognition methods in complex underwater environments, such as weak target features, strong noise interference, significant scale variations, and insufficient robustness, this invention proposes a lightweight target recognition method for underwater sonar images. This method does not rely on a single fixed backbone structure. Instead, it constructs a staged feature enhancement network through shallow feature enhancement, deep channel selection, and multi-scale context modeling at the end layer, thereby improving the accuracy and robustness of sonar image target recognition.

[0007] The technical solution of the present invention is as follows: A lightweight target recognition method for underwater sonar images includes the following steps: First, the underwater sonar image to be identified is acquired and preprocessed, including resizing, normalization, and format conversion, to meet the input requirements of the lightweight target recognition network.

[0008] Secondly, the preprocessed sonar images are input into a lightweight target recognition network. This network performs initial feature extraction through a front-end convolutional module and encodes the features layer by layer through a staged structure.

[0009] Then, in the shallow feature extraction stage, a cross-stage aggregation module (CSA) is introduced. This module enhances the ability to express edge, texture, and local structural information in sonar images through channel splitting, multi-path convolutional feature extraction, channel attention enhancement, feature concatenation, and residual connections.

[0010] Furthermore, in the deep semantic feature extraction stage, an improved inverse residual feature encoding module, EMBConv, is introduced. This module consists of dimensionality-upgrading convolution, depthwise separable convolution, an efficient channel attention module (ECA), and dimensionality-reduced projective convolution. Among them, ECA models local cross-channel interaction relationships through global average pooling and one-dimensional convolution, generating channel weights without explicit channel dimensionality reduction, thereby enhancing the response of target-related channels and suppressing noisy channel responses.

[0011] Finally, a Hybrid Spatial Pyramid Pooling (HSPP) module is introduced before the classification output. This module extracts local salient features through max pooling branches at different scales and combines multi-dilation rate dilated convolution branches and global average pooling branches to obtain global contextual semantic information. The features enhanced by HSPP are input into the classification module, and the final output is the target category recognition result of the sonar image.

[0012] Preferably, the aforementioned cross-stage aggregation module CSA includes channel splitting, multi-path convolutional feature extraction, channel attention enhancement, feature concatenation, convolutional projection, and residual connection processes, which are used to enhance the local structural representation ability of shallow features.

[0013] Preferably, the improved inverse residual feature encoding module EMBConv includes an up-dimensional convolution, a depthwise separable convolution, an efficient channel attention module (ECA), and a down-dimensional projection convolution, wherein the efficient channel attention module (ECA) is used to model cross-channel dependencies without explicit channel dimensionality reduction.

[0014] Preferably, the Hybrid Spatial Pyramid Pooling (HSPP) module includes local multi-scale pooling, feature dimensionality reduction, multi-dilation rate dilated convolution, and global average pooling processes to enhance the network's adaptability to multi-scale sonar targets and complex background reverberation.

[0015] Preferably, the lightweight target recognition network includes a front-end convolutional module, a shallow feature enhancement module, a deep semantic encoding module, a multi-scale context enhancement module, and a classification output module.

[0016] The present invention also provides a lightweight target recognition system for underwater sonar images, comprising: The sonar image input module is used to receive underwater sonar images to be identified. The image preprocessing module is used to preprocess sonar images; The feature extraction module is used to perform the feature extraction and enhancement processes in the above method steps; The target recognition module is used to output the target category recognition results of sonar images.

[0017] In addition, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described lightweight target recognition method for underwater sonar images.

[0018] Compared with the prior art, the beneficial effects of the present invention are: (1) The present invention enhances the shallow local structure modeling capability through the CSA module, enabling the model to retain target edge, texture and local structure information more fully in the high resolution feature map, which is beneficial to improving the target recognition accuracy in low signal-to-noise ratio sonar images.

[0019] (2) This invention introduces an efficient channel attention mechanism through the EMBConv module, which avoids the information bottleneck caused by channel compression, reduces model parameter overhead, and enhances the ability to preserve the echo characteristics of weak targets.

[0020] (3) The present invention integrates local multi-scale response and global contextual semantic information through the HSPP module, enabling the model to better adapt to changes in the scale of underwater targets and complex background reverberation, thereby improving recognition robustness.

[0021] (4) The present invention has a simple overall structure and a clear reasoning path. It can maintain high recognition accuracy while taking into account the lightweight model and convenient deployment. It is suitable for deployment in underwater robots, unmanned underwater vehicles, shore-based sonar image processing systems and edge computing platforms. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall structure of the lightweight target recognition method for underwater sonar images according to the present invention.

[0023] Figure 2 This is a schematic diagram of the Efficient Channel Attention Module (ECA) structure in this invention.

[0024] Figure 3 This is a schematic diagram of the Hybrid Spatial Pyramid Pooling Module (HSPP) structure in this invention. Detailed Implementation

[0025] The application of this invention in underwater sonar target identification will be described in detail below with reference to the accompanying drawings in the embodiments of the present invention. It should be noted that the embodiments described are only a part of the present invention, and other implementations obtained by those skilled in the art based on these embodiments without inventive effort are also within the protection scope of the present invention.

[0026] This invention proposes a lightweight target recognition method for underwater sonar images. The method first acquires the underwater sonar image to be recognized and then performs preprocessing operations such as resizing, normalization, and format conversion to meet the network input requirements. In one embodiment, the input image size can be set to 224×224, which is then input into the lightweight target recognition network.

[0027] like Figure 1 As shown, the lightweight target recognition network mainly includes a front-end convolutional module, a shallow feature enhancement module, a deep semantic coding module, a multi-scale context enhancement module, and a classification output module. The front-end convolutional module is used to perform initial feature extraction and downsampling on the input sonar image to obtain a basic feature map. Subsequently, the network enters a staged feature extraction process.

[0028] In the shallow feature extraction stage, such as Figure 1As shown, this invention introduces a cross-stage aggregation module (CSA). The CSA module first divides the input features along the channel dimension into an identity mapping stream and an enhancement processing stream. The enhancement processing stream extracts local spatial features through convolution operations and recalibrates the channel weights of the convolution output using an efficient channel attention module (ECA) to highlight effective target responses and suppress noisy channels. Then, the identity mapping path, convolutional feature path, and attention enhancement path are concatenated, and feature fusion and dimensionality reduction are achieved through convolutional projection. Finally, the fused features are summed with the residuals of the bypass branches and passed through an activation function to obtain the CSA module output. Through this design, the CSA module can enhance the edge, texture, and local structure representation capabilities of sonar images.

[0029] In the deep semantic encoding stage, such as Figure 1 As shown, this invention employs the EMBConv module for feature extraction. The EMBConv module includes dimensionality-increasing convolution, depthwise separable convolution, an efficient channel attention module (ECA), and dimensionality-reduced projection convolution. Specifically, this module first expands the feature channel representation space through dimensionality-increasing convolution, then extracts spatial features through depthwise separable convolution, subsequently performs channel recalibration using ECA, and finally forms stable output features through dimensionality-reduced projection convolution. Compared with traditional channel attention structures, ECA does not perform explicit channel compression, but instead achieves local cross-channel interaction through global average pooling and one-dimensional convolution. Therefore, it can preserve weak target echo features in sonar images and suppress invalid background responses with lower parameter overhead.

[0030] Before categorizing and outputting, such as Figure 3 As shown, this invention further introduces a Hybrid Spatial Pyramid Pooling (HSPP) module. The HSPP module first feeds the input features into max-pooling branches with different kernel sizes to obtain salient response features at different local scales. Then, the original input features are concatenated with the outputs of each pooling branch, and feature dimensionality reduction is performed through convolution. Next, the features enter a multi-scale context-aware stage, capturing multi-scale contextual information through multiple dilated convolution branches with different dilation rates, while simultaneously extracting overall semantic information of the image using a global average pooling branch. Finally, the outputs of each branch are concatenated and fused again to obtain enhanced multi-scale semantic features.

[0031] Finally, the enhanced feature input is processed by the classification output module. This module integrates channels, aggregates global features, and outputs the probability values ​​for each target category using a classifier. The category with the highest probability is then selected as the final recognition result.

[0032] In this embodiment, the present invention can be trained and validated using publicly available sonar image datasets and their extended datasets. Target categories may include underwater targets such as ship wrecks, plastic bottles, cans, beverage cartons, tires, chains, valves, aircraft wrecks, propellers, hooks, shampoo bottles, upright bottles, and human bodies. Experimental results show that the CSA module, EMBConv module, and HSPP module can all improve the target recognition performance of sonar images. The complete model achieves high recognition accuracy while maintaining a relatively controllable model size, verifying the effectiveness and practicality of the method of the present invention.

[0033] The complete method flow of this invention is as described above. All content not described in detail in this specification belongs to the prior art known to those skilled in the art. Modifications to the embodiments of this invention, equivalent substitutions of technical features, etc., without departing from the spirit and principles of this invention, should be included within the scope of protection of this invention.

Claims

1. A lightweight target recognition method for underwater sonar images, characterized in that, Includes the following steps: (1) Acquire the underwater sonar image to be identified, and perform size adjustment, normalization and preprocessing on the sonar image to obtain sonar image data that meets the network input requirements; (2) Input the preprocessed sonar image into the lightweight target recognition network, and complete the initial feature extraction and downsampling through the front-end convolution module to obtain shallow feature representation; (3) In the shallow feature extraction stage, the cross-stage aggregation module CSA is used to enhance the expression of shallow local structural features through channel splitting, multi-path convolution feature extraction, channel attention enhancement and residual connection. (4) In the deep semantic feature extraction stage, an improved inverse residual feature encoding module EMBConv is used. EMBConv includes dimensionality-increasing convolution, depthwise separable convolution, efficient channel attention module ECA, and dimensionality-reducing projection convolution to enhance the discriminative ability of deep semantic features. (5) Before the deep feature input classification head, the Hybrid Spatial Pyramid Pooling (HSPP) module is introduced to obtain multi-scale contextual semantic features through multi-scale max pooling, feature concatenation, convolution dimensionality reduction, multi-dilation dilated convolution and global average pooling. (6) Input the HSPP-enhanced features into the classification head, and output the sonar image target category recognition result through global feature aggregation and classifier.

2. The method according to claim 1, characterized in that, The cross-stage aggregation module (CSA) includes channel splitting, multi-path convolutional feature extraction, channel attention enhancement, feature concatenation, convolutional projection, and residual connection processes, which are used to enhance the local structural representation ability of shallow features.

3. The method according to claim 1, characterized in that, The improved inverse residual feature encoding module EMBConv includes dimensionality-upgrading convolution, depthwise separable convolution, efficient channel attention module ECA, and dimensionality-reducing projection convolution, wherein the efficient channel attention module ECA is used to model cross-channel dependencies without explicit channel dimensionality reduction.

4. The method according to claim 1, characterized in that, The Hybrid Spatial Pyramid Pooling Module (HSPP) includes local multi-scale pooling, feature dimensionality reduction, multi-dilation rate dilated convolution, and global average pooling processes, which are used to enhance the network's adaptability to multi-scale sonar targets and complex background reverberation.

5. The method according to claim 1, characterized in that, The lightweight target recognition network includes a front-end convolutional module, a shallow feature enhancement module, a deep semantic encoding module, a multi-scale context enhancement module, and a classification output module.

6. A lightweight target recognition system for underwater sonar images, characterized in that, include: The sonar image input module is used to receive underwater sonar images to be identified. The image preprocessing module is used to preprocess the sonar image in step (1) of claim 1; The feature extraction module is used to perform the feature extraction and enhancement process according to any one of claims 1 to 5; The target recognition module is used to output the target category recognition results of sonar images.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the lightweight target recognition method for underwater sonar images as described in any one of claims 1 to 5.

Images