Lightweight Antarctic krill acoustic image detection method based on improved YOLOv8n

Abstract

The application relates to an acoustic image lightweight Euphausia superba detection method based on an improved YOLOv8n, which comprises the following steps: acquiring an acoustic image of Euphausia superba; constructing a target recognition model based on the improved YOLOv8n and training the model; the improved YOLOv8n comprises the following steps: step-by-step feature extraction of an input image by using a RepGhost module, wherein multi-scale features are output after each level of features is enhanced by a C2f-Trans module and a CoECA module; feature fusion by using a BiFPN module, wherein a target detection result is output by a detection head with different detection fields according to the fused features; and inputting the acoustic image into the trained target recognition model to obtain an Euphausia superba group detection result. The application can greatly improve the detection speed, and significantly improve the detection precision and recall rate of underwater weak targets such as Euphausia superba.

Description

Technical Field

[0001] This invention relates to the field of underwater target detection technology, and in particular to a lightweight acoustic image detection method for Antarctic krill based on an improved YOLOv8n. Background Technology

[0002] Underwater acoustic imagery is a core data source for detecting aquatic biological resources such as Antarctic krill, and its detection technology is of great significance for polar fishery resource assessment and intelligent fishing. However, existing underwater acoustic imagery target detection methods face the following problems in practical applications:

[0003] In underwater environments, sound wave propagation is affected by factors such as water scattering, seabed / surface reverberation, and equipment electrical noise, resulting in complex texture noise in the background of sonar images. Traditional target detection models are prone to misidentifying noise as targets or missing real targets.

[0004] The echo signals of small aquatic organisms such as Antarctic krill are weak, resulting in extremely low contrast between the target and the background in sonar images. Traditional target recognition networks rely on local receptive fields to extract features, which makes it difficult to capture the edge and texture details of weak targets, and is prone to missed detection or positioning errors.

[0005] Shipborne equipment is limited by power consumption and size, and has limited computing resources. Furthermore, existing detection models have a large number of parameters and high computational complexity, which cannot meet the needs of real-time monitoring. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a lightweight acoustic image detection method for Antarctic krill based on the improved YOLOv8n, which can significantly improve the detection speed while significantly improving the detection accuracy and recall rate of weak underwater targets such as Antarctic krill.

[0007] The technical solution adopted by this invention to solve its technical problem is: to provide a lightweight acoustic image detection method for Antarctic krill based on improved YOLOv8n, comprising:

[0008] Acoustic images of Antarctic krill were obtained;

[0009] Construct and train a target recognition model based on an improved YOLOv8n; the improved YOLOv8n includes:

[0010] The RepGhost module is used to extract features from the input image step by step. Each level of features is enhanced by the C2f-Trans module and the CoECA module in sequence to output multi-scale features.

[0011] A BiFPN module is introduced to perform multi-scale feature fusion, and the target detection results are output based on the fused features through the detection heads of different detection fields;

[0012] The acoustic images are input into the trained target recognition model to obtain the detection results of Antarctic krill populations.

[0013] Furthermore, the CoECA module includes a first attention branch and a second attention branch for parallel processing of input features. The first attention branch is used to extract key information in the channel dimension and output a channel attention mask, while the second attention branch is used to extract key information in the spatial dimension and output a spatial attention mask. The channel attention mask and the spatial attention mask are fused based on the collaborative weight coefficients, and the fused attention mask is used to enhance the input features and output the enhanced features.

[0014] Furthermore, the first attention branch consists of sequentially connected global average pooling and 1D convolutions, and generates a channel attention mask by activating the Sigmoid function.

[0015] Furthermore, the second attention branch consists of a 3×3 depthwise separable convolution, which generates a spatial attention mask by activation of the Sigmoid function.

[0016] Furthermore, the attention mask is multiplied with the input feature map to achieve feature enhancement.

[0017] Furthermore, during the training of the target recognition model, the RepGhost module includes a first convolutional branch, a second convolutional branch, and a third convolutional branch that perform parallel feature extraction on the input image. The outputs of the three convolutional branches are summed and then activated through a BN layer and a SiLU function. The first convolutional branch is used to extract local spatial features, the second convolutional branch is used to extract inter-channel interaction features, and the third convolutional branch is used to preserve the original features when the number of input and output channels is the same.

[0018] Furthermore, during the inference process of the target recognition model, the first, second, and third convolutional branches are reparameterized into single-path 3×3 convolutions.

[0019] Furthermore, the C2f-Trans module replaces the bottleneck block in the C2f module with the Transformer module.

[0020] Furthermore, the acoustic image is a sonar image generated after analyzing the acoustic data collected by the echo sounder.

[0021] Beneficial effects

[0022] Due to the adoption of the above-mentioned technical solutions, this invention has the following advantages and positive effects compared with the prior art: This invention, through RepGhost reparameterization design, has a small number of model parameters and fast inference speed, which can effectively meet the needs of shipborne real-time detection; with the optimization of CoECA and BiFPN, the accuracy of Antarctic krill swarm detection is improved; in addition, the unique attention weighting strategy effectively suppresses seabed echo noise, enhances the adaptability to low-contrast and blurred targets, and is suitable for complex polar water environments. Attached Figure Description

[0023] Figure 1 This is a structural diagram of the RG-EA-YOLOv8n detection model according to an embodiment of the present invention;

[0024] Figure 2 This is a structural diagram of the RepGhost module according to an embodiment of the present invention;

[0025] Figure 3 This is a structural diagram of the C2f-Trans module according to an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the CoECA mechanism according to an embodiment of the present invention;

[0027] Figure 5 This is a diagram showing the detection results of Antarctic krill populations according to an embodiment of the present invention. Detailed Implementation

[0028] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0029] The embodiments of the present invention relate to a lightweight underwater acoustic image target detection method based on improved YOLOv8, which aims to solve the problems of low detection accuracy and slow speed caused by complex background noise, low target contrast, and limited computing power of shipborne equipment in underwater sonar images.

[0030] This implementation method improves upon YOLOv8n, constructing the RG-EA-YOLOv8n detection model, the model architecture of which is as follows. Figure 1 As shown:

[0031] First, the RepGhost reparameterization module is introduced into the backbone network, which merges the multi-branch structure into a single-path convolution during the inference stage, significantly reducing the amount of computation.

[0032] Secondly, by combining the C2f-Trans module, the Transformer mechanism is used to capture global context information;

[0033] Secondly, a collaborative efficient channel-spatial attention mechanism (CoECA) is designed to collaboratively enhance semantic features and spatial boundaries through a specific weighting strategy. Finally, a BiFPN structure is used to perform bidirectional weighted fusion of multi-scale features.

[0034] The RepGhost module eliminates redundant computation by fusing multi-branch structures into single-path convolutions during the inference phase. The C2f-Trans module is introduced to capture global contextual information using the Transformer mechanism. The CoECA mechanism is proposed to collaboratively enhance semantic features and spatial boundaries through a specific weighting strategy. Finally, BiFPN is used to achieve efficient bidirectional feature pyramid fusion, enabling efficient and accurate detection of Antarctic krill populations in complex underwater environments.

[0035] More specifically, acoustic images used for target recognition can be obtained through the following methods:

[0036] The raw data (.Raw files) collected by the scientific echo sounder are parsed and converted into visualized sonar images;

[0037] The images are uniformly scaled to 640×640 pixels and normalized to construct a standard input tensor.

[0038] The preprocessed data is input into the backbone network. The RepGhost module is used for downsampling, which is equivalent to a 3×3 convolution during inference. The C2f-Trans module is used to capture long-range semantic dependencies through a multi-head self-attention mechanism (MHSA).

[0039] like Figure 2 As shown, the RepGhost module employs a multi-branch parallel topology during the training phase to improve feature representation capabilities, specifically including:

[0040] The 3×3 convolutional branch is used to extract local spatial features and capture the edge and texture information of the target.

[0041] A 1×1 convolutional branch is used to realize linear feature interaction between channels, enhancing the information flow across channels;

[0042] The identity branch is enabled only when the module step size is 1 and the number of input and output channels are the same. It is used to directly pass the input features to subsequent layers, effectively reducing information loss in deep networks.

[0043] During training, the outputs of the three branches are element-wise summed and then sequentially passed through a batch normalization (BN) layer and the SiLU activation function. During inference, the multi-branch weights are fused into an equivalent single-path 3×3 convolutional layer using structural reparameterization, thereby significantly reducing the computational overhead and inference latency of the model on shipboard equipment while maintaining feature extraction accuracy.

[0044] The RepGhost module, as the underlying computational unit of the backbone network, employs a structural reparameterization strategy. During training, it uses a complex multi-branch structure to enhance feature representation. During inference deployment, it merges all branch weights into a single 3×3 convolutional kernel. This achieves computational compression and inference acceleration without sacrificing accuracy, addressing the stringent speed requirements of low-computing-power shipboard devices. A C2f-Trans module is embedded on this lightweight skeleton to capture long-range global contextual information that is difficult for traditional convolutions to perceive. Figure 3 As shown, this module introduces the Transformer's multi-head self-attention mechanism (MHSA) into the C2f structure, enabling the network to establish correlations between distant pixels on deep feature maps, effectively solving the limitations of local features caused by the dispersion of krill targets and complex backgrounds in sonar images.

[0045] like Figure 4 As shown, Attention Enhancement (CoECA) is used to apply a collaborative attention mechanism to the backbone output feature map. It generates an enhanced feature map by performing global average pooling and 1D convolution (channel branch) and 3×3 depthwise separable convolution (spatial branch) in parallel, and then multiplying them by the collaborative weight coefficients (α=0.6, β=0.4) and fusing them.

[0046] More specifically, the input to the CoECA module is the feature map output by the backbone network. Key information in the channel dimension and spatial dimension is extracted through parallel channel branches and spatial branches, respectively. Finally, the enhanced feature map is generated by fusion through collaborative weights.

[0047] The goal of the channel attention branch is to capture the dependencies between channels and focus on the feature channels that are more critical to the task. Its operation process is as follows:

[0048] 1) Global Average Pooling (GAP):

[0049] Global average pooling is performed on the input feature map along the spatial dimension to obtain a vector of (B, C, 1, 1). This operation compresses spatial information into channel-dimensional statistics while preserving the global context information of each channel.

[0050] 2) 1D convolution (1×3Conv1D):

[0051] The pooled vector is adjusted to a sequence of (B,1,C), and local dependencies between channels are captured by a 1×3 1D convolution. 1D convolution has fewer parameters than fully connected layers, which meets the requirements of lightweight design.

[0052] 3) Sigmoid normalization:

[0053] After the 1D convolution output is adjusted back to the (B,C,1,1) shape, a channel attention mask mask_c is generated by the Sigmoid activation function. The magnitude of the mask value represents the importance of the corresponding channel.

[0054] The goal of the spatial attention branch is to locate the spatial regions in the feature map that are key to object detection. Its operation process is as follows:

[0055] 1) 3×3 depthwise separable convolution:

[0056] The input feature map is processed by a 3×3 depthwise separable convolution. Compared with a regular 3×3 convolution (with C×1×3×3 parameters), the depthwise separable convolution has only C×3×3+C×1×1×1 parameters, which is about 1 / 3 of the computation of a regular 3×3 convolution, further reducing the module complexity.

[0057] 2) Sigmoid normalization

[0058] The convolution output is activated by Sigmoid to generate a spatial attention mask mask_s (size (B,1,H,W)). The larger the mask value, the higher the importance of the corresponding spatial location.

[0059] The masks `mask_c` and `mask_s` generated by the channel branch and spatial branch respectively indicate which channels or locations are more important in the channel and spatial dimensions. However, single-dimensional attention may miss crucial information in the other dimension; for example, a strong channel response may correspond to a background region, while a strong spatial response may correspond to a low-discrimination channel. Therefore, the CoECA module fuses the two masks using collaborative weighting coefficients.

[0060]

[0061] Where α=0.6 and β=0.4.

[0062] The input feature map is multiplied element-wise with the fusion mask to achieve dual enhancement of key channels and key spatial locations.

[0063] The CoECA (Enhanced Attention) module combines the advantages of coordinate attention and efficient channel attention (ECA), improving the model's target response capability and global feature modeling efficiency. CoECA extracts global context in the height and width dimensions separately through adaptive average pooling, generating independent attention weights for low-cost fusion. While retaining the positional sensitivity of coordinate attention, it employs channel adaptive calibration of ECA, reducing the number of parameters and making it suitable for real-time detection scenarios. Simultaneously, CoECA enhances response accuracy for small targets and complex backgrounds. After application to the BiFPN neckline, it improves the globality of feature fusion, reduces overall model complexity, and increases inference speed, achieving a balance between real-time performance and accuracy while maintaining high mAP.

[0064] Example 1: Study of sonar data collected by the Antarctic krill research vessel "Deep Blue"

[0065] The EK80 device used in this embodiment operates at a frequency of 38 kHz. Data acquisition took place from January 23, 2023 to June 16, 2023, in Antarctic krill fishing area 48. The acoustic data detection range was 0-1100 m, and the data included krill information and seabed depth information, which provided support for further research on seabed identification and depth calculation methods. The EK80 single-ping acoustic data consisted of 75,765 echo points, achieving a detection accuracy of 1.5 × 10⁻² m, meeting the data accuracy requirements for krill identification.

[0066] A total of 1949 preprocessed sonar images were used, and the dataset was randomly divided into three parts in an 8:1:1 ratio: 1560 images for training, 194 images for testing, and 195 images for validation.

[0067] Design a lightweight deep detection network with YOLOv8n as the main architecture. The RepGhost module and C2f-Trans module are integrated in the backbone network, and the CoECA attention mechanism and BiFPN structure are integrated in the neck network.

[0068] The network was trained end-to-end using a labeled Antarctic krill sonar image dataset. The RepGhost multi-branch structure was utilized to capture rich features, and network weights were updated via backpropagation. The model was trained using Python 3.12 and PyCharm 2025.1.1. Experiments showed that the accuracy and recall of the model stabilized after 70 training epochs. Therefore, to further optimize model performance, the training epochs were ultimately set to 100. The network model training environment parameters selected to further simulate the computing power of a shipboard industrial control computer are shown in Table 1 below.

[0069] Table 1

[0070] Pytorch 2.5.1 Python 3.12 Cuda 12.1 GPU NVIDIA-RTX3050 Laptop CPU 12th Gen Intel(R) Core(TM) i5-12400 operating system Windows 11

[0071] During the inference phase, an operator fusion strategy is used to convert the multi-branch structure of RepGhost into a single-path convolution, compressing model memory and accelerating inference. Based on the features of the input sonar image, the target location bounding box, category, and confidence score of the Antarctic krill swarm are output.

[0072] The RG-EA-YOLOv8n network includes the following core modules:

[0073] a. Lightweight Backbone: The RepGhost module is used to replace the traditional convolutional layer for feature extraction, reducing the number of parameters; the C2f-Trans module is used to replace some bottleneck layers, and long-distance semantic information is captured through a global self-attention mechanism.

[0074] b. Co-Attention Layer (CoECA): A CoECA module is embedded at the main feature output end, and a dual-branch parallel structure is used to extract channel and spatial features simultaneously, enhancing the ability to perceive low-contrast krill targets.

[0075] c. Multi-scale fusion layer (Neck): The original PANet structure is replaced with BiFPN (Bidirectional Feature Pyramid Network) to achieve bidirectional cross-scale fusion of features from top to bottom and from bottom to top.

[0076] d. Detection Head: Receives the fused feature map and outputs the final target detection result.

[0077] The module employs the following specific computational mechanisms, feature enhancement, and optimization strategies:

[0078] a. RepGhost reparameterization mechanism: During training, it employs a multi-branch topology (3×3 convolution, 1×1 convolution, identity mapping); during inference, it merges them into a single 3×3 convolution kernel.

[0079] b. CoECA collaborative weighting mechanism: Explicit collaborative weight coefficients are introduced to fuse channel branches and spatial branches. Weights are adaptively learned through convolutional layers to prioritize the enhancement of semantic features and suppress environmental noise.

[0080] c. C2f-Trans global awareness strategy: Embed TransformerBlock in C2f structure and use multi-head self-attention mechanism (MHSA) to interact on flattened feature maps to solve the problem of limited receptive field of local convolution.

[0081] d. BiFPN weighted fusion strategy: Introduce a simple additive fusion mechanism to perform non-equal weighted combination of input feature maps of different resolutions, thereby improving the model's adaptability to krill populations of different densities.

[0082] Test results as follows Figure 5 As shown, the RG-EA-YOLOv8n model in this embodiment achieves significant performance gains compared to the benchmark YOLOv8n model on the Antarctic krill acoustic dataset, with a 10% improvement in precision, a 7% improvement in recall, and a 6.7% improvement in mAP@0.5. At the same time, the inference speed is improved by about 25%. While maintaining the advantage of lightweight design, it outperforms the comparative models such as YOLOv8-BiFPN in both quantitative and qualitative analysis.

[0083] Compared with existing technologies, this invention offers faster processing speed and higher accuracy. Through RepGhost reparameterization, the model parameter count is controlled to around 3.4M, resulting in an inference speed (FPS) improvement of approximately 17%-30% compared to the original YOLOv8n. The processing time per image is only about 37ms, meeting the requirements for real-time shipboard detection. High accuracy and good stability: With optimizations using CoECA and BiFPN, the accuracy reaches 92.61% on the Antarctic krill dataset, and mAP@0.5 reaches 92.37%, significantly outperforming commonly used models such as YOLOv8s and YOLOv8-BiFPN. Wide applicability and strong noise resistance: The unique attention-weighted strategy effectively suppresses seabed echo noise, enhancing adaptability to low-contrast and blurred targets, making it suitable for complex polar aquatic environments.

Claims

1. A lightweight acoustic image detection method for Antarctic krill based on an improved YOLOv8n, characterized in that, include: Acoustic images of Antarctic krill were obtained; Construct and train an object recognition model based on the improved YOLOv8n; The improved YOLOv8n includes: The RepGhost module is used to extract features from the input image step by step. Each level of features is enhanced by the C2f-Trans module and the CoECA module in turn to output multi-scale features. A BiFPN module is introduced to perform multi-scale feature fusion, and the target detection results are output based on the fused features through the detection heads of different detection fields; The acoustic images are input into the trained target recognition model to obtain the detection results of Antarctic krill populations.

2. The method for detecting Antarctic krill according to claim 1, characterized in that, The CoECA module includes a first attention branch and a second attention branch for parallel processing of input features. The first attention branch is used to extract key information in the channel dimension and output a channel attention mask, while the second attention branch is used to extract key information in the spatial dimension and output a spatial attention mask. The channel attention mask and the spatial attention mask are fused based on the collaborative weight coefficients, and the fused attention mask is used to enhance the input features and output the enhanced features.

3. The method for detecting Antarctic krill according to claim 2, characterized in that, The first attention branch consists of sequentially connected global average pooling and 1D convolutions, and generates a channel attention mask by activating the Sigmoid function.

4. The method for detecting Antarctic krill according to claim 2, characterized in that, The second attention branch consists of a 3×3 depthwise separable convolution, which generates a spatial attention mask by activating the sigmoid function.

5. The method for detecting Antarctic krill according to claim 2, characterized in that, Feature enhancement is achieved by multiplying the fused attention mask with the input feature map.

6. The method for detecting Antarctic krill according to claim 1, characterized in that, During the training of the target recognition model, the RepGhost module includes a first convolutional branch, a second convolutional branch, and a third convolutional branch that perform parallel feature extraction on the input image. The outputs of the three convolutional branches are summed and then activated by a BN layer and a SiLU function. The first convolutional branch is used to extract local spatial features, the second convolutional branch is used to extract inter-channel interaction features, and the third convolutional branch is used to preserve the original features when the number of input and output channels is the same.

7. The method for detecting Antarctic krill according to claim 6, characterized in that, During the inference process of the target recognition model, the first, second, and third convolutional branches are reparameterized into single-path 3×3 convolutions.

8. The method for detecting Antarctic krill according to claim 1, characterized in that, The C2f-Trans module replaces the bottleneck block in the C2f module with the Transformer module.

9. The target detection method according to claim 1, characterized in that, The acoustic image is a sonar image generated after analyzing the acoustic data collected by the echo sounder.

Images