Enhanced biological image detection method for complex underwater environment

Abstract

The invention relates to an enhanced biological image detection method for a complex underwater environment, and belongs to the technical field of underwater biological image detection.The method comprises the steps that an obtained underwater image is input into an EFE-UODNet enhanced biological detection network, an underwater biological detection result is output, the EFE-UODNet enhanced biological detection network is improved based on a YOLOv8 network, and the EFE-UODNet enhanced biological detection network is used for detecting the underwater biological image. In a backbone network, an EGCGF module and a GCGF module are utilized to replace a C2F module to perform efficient multi-scale feature extraction, an AM-FPN network is utilized to replace an original neck network to fully utilize context information of an underwater environment to perform feature fusion, and a dynamic detection head is utilized to replace an original detection head to output a detection result through a specific multi-scale target detection mechanism. Compared with the prior art, the method has the advantages that the detection precision of the marine organisms in the complex underwater environment is improved, and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater biological image detection, and in particular to an enhanced biological image detection method for complex underwater environments. BACKGROUND

[0002] The ocean accounts for 71% of the Earth's surface, and its resources have both ecological and economic value. Rational development is crucial for sustainable development. Autonomous underwater vehicles (AUVs) combined with computer vision technology can efficiently perform marine detection tasks, reducing human intervention, reducing operating costs, and improving operational efficiency, which is of great significance for monitoring marine biodiversity and distribution patterns, and maintaining ecological balance and resource management. However, underwater target detection faces significant challenges: first, the complex underwater environment leads to poor image quality, suspended solids, light attenuation, and scattering factors that blur the color and shape features of the target; second, benthic organisms are unevenly distributed, and similar targets are densely packed and similar in features, making it difficult for detectors to accurately divide boundaries; third, the size of the target varies greatly, with small objects accounting for a small proportion and having few features, which can easily lose information during downsampling, affecting detection accuracy. Current general algorithms perform well on land, but need to be optimized for underwater scenarios to improve recognition capabilities in complex environments.

[0003] In recent years, underwater target detection technology has been widely used in marine engineering and autonomous underwater vehicles (AUVs), and has received considerable attention. Liu et al. proposed a multi-scale feature pyramid network based on deep learning, further optimizing the feature pyramid network to enhance the feature extraction capability related to small targets. This network is used for automatic detection of underwater organisms, providing effective technical support for automated breeding and harvesting of marine aquaculture. However, these methods still face challenges when the class distribution of the dataset is uneven. Li et al. proposed an improved algorithm based on Faster-RCNN, which improves the accuracy of fish detection by optimizing the generation method of candidate regions. The spatial pyramid pooling network (SPP-Net) accelerates the target detection process by introducing a spatial pyramid pooling structure, thereby improving feature utilization. However, SPP-Net still relies on candidate regions, which limits its application in underwater environments. Chinese patent CN118097393A discloses an underwater biological target detection method based on a YOLOv8 model, which fuses channel attention mechanisms and spatial attention mechanisms to obtain a residual attention mechanism; embeds a bidirectional feature pyramid network and a residual attention mechanism into the YOLOv8 model to obtain an optimized YOLOv8 model for underwater biological detection. However, this method lacks attention to small-sized underwater organisms in the detection part of the network. Therefore, there is currently a lack of a biological detection method that can handle poor image quality, significant scale differences, and uneven distribution of organisms, while balancing accuracy and robustness in complex environments. SUMMARY

[0004] The purpose of the present application is to provide an enhanced biological image detection method for complex underwater environments, and the constructed EFE-UODNet enhanced biological detection network is based on the YOLOv8 architecture for multi-scale fusion enhancement, effectively reducing the small target missing detection rate and target positioning deviation, and can better detect blurred, dense and small size objects, thereby improving the marine biological detection precision in complex underwater environments.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] An enhanced biological image detection method for complex underwater environments, which inputs the obtained underwater image into an EFE-UODNet enhanced biological detection network and outputs underwater biological detection results, wherein the EFE-UODNet enhanced biological detection network is improved based on the YOLOv8 network, the C2F module in the backbone network is replaced by the EGCGF module and the GCGF module for efficient multi-scale feature extraction, the original neck network is replaced by the AM-FPN network to fully utilize the context information of the underwater environment for feature fusion, and the original detection head is replaced by a dynamic detection head to output the detection results through its unique multi-scale target detection mechanism.

[0007] The EGCGF module includes five deep separable convolution modules with different sizes of large convolution kernels and a GCGF module, two deep separable convolution modules in it form a deep separable convolution processing channel, a total of four separable convolution processing channels and a GCGF module processing channel are constructed, and the input is processed respectively, and the five outputs are spliced to obtain the output of the EGCGF module, wherein the deep separable convolution module is composed of a depth convolution and a point-by-point convolution, the depth convolution is used for extracting spatial features, and the point-by-point convolution is used for extracting channel features.

[0008] The GCGF module performs convolution processing on the input features, then divides the feature map to obtain a first feature map and a second feature map, the output of the second feature map after processing by N series of EB modules is spliced with the first feature map and the second feature map, and then output after a layer of convolution.

[0009] The EB module is obtained by connecting a dynamic convolution and a VSSBlock module in series.

[0010] The EGCGF module performs the following operations on the input feature map F:

[0011]

[0012] Conv c1×c2represents a convolution operation with a convolution kernel size of c1xc2 on the feature map, that is, represents a two-dimensional convolution operation, that is, a convolution calculation, BN represents batch normalization, Silu represents an activation function, c1 and c2 respectively represent the width and height of the convolution kernel; N is the number of EB modules in series; Split represents a splitting operation; EB represents the operation of the EB module, EB(F) = VSSBlock(DynamicConv 3×3 (F))+F, DynamicConv 3×3 represents a dynamic convolution operation with a convolution kernel size of 3x3 on the feature map, VSSBlock represents the operation of the VSSBlock module; GCGF(F) represents the output of the GCGF module; DWConv c1×c2 represents a depth separable convolution operation with a convolution kernel size of c1xc2 on the feature map; Concat represents a feature map splicing function.

[0013] The backbone network of the EFE-UODNet enhanced biological detection network comprises two convolution modules connected in sequence, a first EGCGF module, a convolution module, a second EGCGF module, a convolution module, a third EGCGF module, a convolution module, a GCGF module and a SPPF module, wherein the first EGCGF module, the second EGCGF module, the third EGCGF module and the SPPF module output first-level features, second-level features, third-level features and fourth-level features respectively, and input an AM-FPN network.

[0014] The AM-FPN network sets four feature processing levels, receives the first to fourth level features output by the backbone network to capture micro target details, small target details, medium target details and large target details respectively, and comprises three feature fusion modules and three feature extraction modules. The first level features input a first feature fusion module, the second level features input a second feature fusion module, the third level features input a third feature fusion module, and the fourth level features input the third feature fusion module and a third feature extraction module. The output of the third feature fusion module is input to a second feature extraction module, the outputs of the first and second feature fusion modules are input to a first feature extraction module, and the outputs of the first to third feature extraction modules and the input of the first feature extraction module constitute four outputs of the AM-FPN network, which are input to corresponding four dynamic detection heads respectively.

[0015] The feature fusion module is composed of a FFMC module and a C2F module in cascade, and the feature extraction module is composed of a Conv layer, a Concat layer and a C2F module in cascade.

[0016] The FFMC module comprises two inputs: relatively low-level features and relatively high-level features, and the FFMC module sequentially performs transpose convolution operation, bilinear interpolation processing, CBAM module attention processing and two-dimensional convolution operation on the relatively high-level features, and then elementarily fuses the relatively low-level features to obtain the output of the FFMC module.

[0017] The FFMC module performs the following operations on the input features:

[0018] f out = Conv2d (CBAM (BI (TC (f high ))) + f low

[0019] Wherein, f low ∈R 2H×2W×C represents the relatively low-level features, f high ∈R H×W×C represents the relatively high-level features, TC represents the transpose convolution operation, BI represents the bilinear interpolation processing, CBAM represents the CBAM module attention processing operation, Conv2d represents the two-dimensional convolution operation, f out ∈R 2H×2W×C represents the features output by the FFMC module.

[0020] Compared with the prior art, the present application has the following beneficial effects:

[0021] (1) The present application proposes an enhanced underwater biological detection network EFE-UODNet in a complex environment, aiming to improve the fine-grained feature extraction capability of biological in a complex underwater scene. This structure shows stronger feature discrimination ability and expression accuracy when dealing with image quality difference, significant scale difference and uneven biological distribution.

[0022] (2) The present application designs a feature extraction module named EGCGF, which maintains linear complexity while having the advantages of global receptive field and dynamic weighting, thereby significantly improving the feature extraction efficiency. They can effectively compensate for the loss of object edge and texture details in underwater scenes, fully capture and describe the local features of underwater environment, so that the network can more efficiently learn multi-scale feature information, and further improve the discrimination ability of underwater target organisms and background.

[0023] (3) The adaptive multi-scale feature pyramid network (AM-FPN) of the present application is used for feature extraction and fusion to fully utilize the context information of underwater environment. This improvement enables the network to effectively extract feature information of large, medium, small and even tiny targets from multi-layer feature maps, thereby further improving the precision and robustness of underwater biological detection.

[0024] (4) On the widely used complex underwater environment biological image dataset DUO, the EFE-UODNet network of the present invention has more advantages in feature extraction and can effectively perform the task of detecting organisms in complex underwater environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flowchart of the application of EFE-UODNet enhanced biological detection network in the present invention;

[0026] Figure 2 This is the architecture diagram of the EFE-UODNet enhanced biodetection network in the present invention;

[0027] Figure 3 Schematic diagram of the structure of the EGCGF module in the present invention;

[0028] Figure 4 Schematic diagram of the structure of the AM-FPN network in the present invention;

[0029] Figure 5 Schematic diagram of the structure of the FFMC module in the present invention;

[0030] Figure 6 This is a visualization result of an ablation experiment in one embodiment of the present invention;

[0031] Figure 7 This is a comparison result of visual detection between one embodiment of the present invention and other algorithms. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts should fall within the scope of protection of the present invention.

[0033] This embodiment provides an enhanced biological image detection method for complex underwater environments. Figure 1 As shown in , this method inputs the acquired underwater image into the EFE-UODNet enhanced biological detection network and outputs the underwater biological detection results. Figure 2As shown, the EFE-UODNet enhanced biological detection network is improved based on the YOLOv8 network. On the basis of inheriting the traditional feature extraction and fusion framework, the feature expression ability and multi-scale adaptability are systematically enhanced. The network as a whole still follows the feature progressive processing paradigm from the bottom to the top, but in the backbone network, the EGCGF module and the GCGF module are used to replace the C2F module for efficient multi-scale feature extraction, the AM-FPN network is used to replace the original neck network to fully utilize the context information of the underwater environment for feature fusion, and the dynamic detection head is used to replace the original detection head to output the detection results through its unique multi-scale target detection mechanism.

[0034] Firstly, the enhanced global context guided feature module (EGCGF) and the global context guided feature module (GCGF) are embedded in the backbone network for feature extraction. These modules, while maintaining linear complexity, combine the advantages of global receptive field and dynamic weighting, significantly improving feature extraction efficiency. They can effectively compensate for the loss of object edges and texture details in underwater scenes, comprehensively capture and describe local features in underwater environments, enabling the network to more efficiently learn multi-scale feature information and thus improve the ability to distinguish underwater target organisms and backgrounds.

[0035] Secondly, the adaptive multi-scale feature pyramid network (AM-FPN) is designed for feature extraction and fusion to fully utilize the context information of the underwater environment. This module integrates channels and spatial information at different depths, and combines convolution and deconvolution operations to ensure consistency in feature dimensions. Meanwhile, the high-level feature pyramid network is applied to the detection head structure to optimize classification and detection performance. This improvement enables the network to effectively extract feature information of large, medium, small, and even tiny targets from multi-layer feature maps, further improving the precision and robustness of underwater biological detection.

[0036] Finally, an advanced dynamic detection head structure (DyHead) is introduced and applied to the field of underwater biological detection. This module takes the feature maps output by the AM-FPN as input and significantly improves the performance of small target detection in complex underwater environments through its unique multi-scale target detection mechanism.

[0037] The EGCGF module is integrated into the YOLOv8 backbone network, which aims to effectively extract and analyze the features of objects in the underwater environment, thereby more efficiently extracting features from underwater optical images and obtaining shallow and deep feature maps. Shallow feature maps have high-resolution feature projections, but the extracted feature information is relatively basic and limited. In contrast, deep feature maps have lower resolution but contain more complex and advanced feature information. The combination of shallow and deep feature maps is crucial for the model to analyze the features of underwater objects.

[0038] AsFigure 3 As shown, the EGCGF module includes a depth separable convolution module with five large convolution kernels of different sizes (the convolution kernel sizes are 13x13, 15x15, 17x17, 19x19, and 21x21, respectively) and a GCGF module, two depth separable convolution modules in the EGCGF module form a depth separable convolution processing channel, a total of four depth separable convolution processing channels are constructed (the convolution kernels of the depth separable convolution modules in the first channel are 13x13 and 15x15, the convolution kernels of the depth separable convolution modules in the second channel are 15x15 and 17x17, the convolution kernels of the depth separable convolution modules in the third channel are 17x17 and 19x19, and the convolution kernels of the depth separable convolution modules in the fourth channel are 19x19 and 21x21), and a GCGF module processing channel, which respectively processes the input and splices the five outputs to obtain the output of the EGCGF module. Using several large convolution kernels instead of multiple small convolution kernels can more effectively generate larger effective receptive fields and significantly improve the performance of the convolutional neural network (CNN). Among them, the depth separable convolution module is composed of depth convolution and pointwise convolution, the depth convolution is used to extract spatial features, and the pointwise convolution is used to extract channel features. By combining these advantages, multiple branches can effectively extract the features of marine organisms in complex environments. Therefore, the EGCGF module can comprehensively learn the shape, texture, and color features of underwater objects and extract background information from images. It can more effectively extract and analyze features from low-quality underwater optical images, significantly improve the performance of the model in extracting underwater biological features, and further enhance its overall detection capability.

[0039] After the GCGF module convolves the input features, the feature map is segmented to obtain a first feature map and a second feature map, the output of the second feature map after being processed by N EB modules in series is spliced with the first feature map and the second feature map, and then output after a layer of convolution. Among them, the EB module is obtained by concatenating a dynamic convolution and a VSSBlock module and then performing residual connection.

[0040] The EGCGF module performs the following operations on the input feature map F:

[0041]

[0042]

[0043] Among them, Conv c1×c2 represents a convolution operation on the feature map with a convolution kernel size of c1x c2, that is represents a two-dimensional convolution operation, that is, a convolution calculation, BN represents batch normalization, Silu represents an activation function, c1 and c2 represent the width and height of a convolution kernel respectively; N is the number of EB modules in series; Split represents a split operation; EB represents the operation of an EB module, EB(F) = VSSBlock(DynamicConv 3×3 (F))+F, DynamicConv 3×3 represents a dynamic convolution operation with a convolution kernel size of 3x3 on a feature map, and VSSBlock represents the operation of a VSSBlock module; GCGF(F) represents the output of a GCGF module; DWConv c1×c2 represents a depth separable convolution operation with a convolution kernel size of c1xc2 on a feature map; and Concat represents a feature map splicing function.

[0044] As shown in Figure 2 , the backbone network of the EFE-UODNet enhanced biological detection network includes two convolution modules connected in sequence, a first EGCGF module, a convolution module, a second EGCGF module, a convolution module, a third EGCGF module, a convolution module, a GCGF module, and a SPPF module. The first EGCGF module, the second EGCGF module, the third EGCGF module, and the SPPF module output a first-level feature, a second-level feature, a third-level feature, and a fourth-level feature respectively, which are input into the AM-FPN network.

[0045] The AM-FPN network sets four feature processing levels to receive the first to fourth level features output by the backbone network to capture micro target details, small target details, medium target details, and large target details respectively. As shown in Figure 4 , the AM-FPN network includes three feature fusion modules and three feature extraction modules. The first-level feature is input into the first feature fusion module, the second-level feature is input into the second feature fusion module, the third-level feature is input into the third feature fusion module, and the fourth-level feature is input into the third feature fusion module and the third feature extraction module. The output of the third feature fusion module is input into the second feature extraction module, and the outputs of the first and second feature fusion modules are input into the first feature extraction module. The outputs of the first to third feature extraction modules and the input of the first feature extraction module constitute four outputs of the AM-FPN network, which are input into corresponding four dynamic detection heads respectively. The feature fusion module is composed of a FFMC module and a C2F module in cascade, and the feature extraction module is composed of a Conv layer, a Concat layer, and a C2F module in cascade.

[0046] The FFMC module is the core component of the AM-FPN network. In the underwater organism detection task, the complexity of the environment and the diversity of targets pose a huge challenge to the detection ability of the model. These complexities are mainly reflected in the multi-scale characteristics of underwater organisms: different species have different sizes and shapes, and even the same species may change its appearance under different conditions. In addition, the uneven lighting, turbid water quality and background noise in the underwater environment further aggravate the difficulty of the detection task. In order to cope with the multi-scale challenges inherent in the underwater organism dataset, the present invention designs a convolutional feature fusion module (FFMC). At the same time, based on the Neck part of YOLOv8, an advanced multi-scale feature pyramid network (AM-FPN) is designed.

[0047] Due to multiple downsampling operations, the spatial resolution of high-level features differs significantly from that of low-level features. Directly fusing these two features will face the problem of spatial dimensional mismatch. To better integrate high-level features with low-level features in terms of spatial scale, the high-level features need to be upsampled to increase their spatial resolution and bring them closer to the low-level features in terms of spatial dimension, facilitating subsequent operations such as element-by-element fusion. To address this issue, transposed convolution is employed. It can learn appropriate convolution kernel parameters. While improving the spatial resolution of the feature map, it can also adjust and generate new feature values ​​based on the input high-level semantic features, preserving and adjusting the semantic information contained in the high-level features to a certain extent. Compared to simpler upsampling methods such as interpolation, transposed convolution has stronger learning capabilities and can better adapt to feature upsampling requirements under different data distributions, laying a solid foundation for subsequent more accurate feature fusion. However, relying solely on transposed convolution may not be able to accurately adjust the spatial resolution of high-level features to be exactly the same as that of low-level features. To this end, the present invention introduces bilinear interpolation. As a classic image interpolation method, bilinear interpolation calculates new pixel values ​​based on the weighted average of surrounding pixels. In the process of adjusting the size of the feature map, it can more finely adjust the spatial size of the high-level features on the basis of maintaining feature continuity and smoothness, so that it matches the low-level features and ensures that the two can be accurately aligned in the spatial dimension, providing a good prerequisite for subsequent fusion operations and reducing the fusion error caused by spatial size differences. Based on this, Figure 5 As shown in the figure, the FFMC module includes two inputs: relatively low-level features and relatively high-level features. The FFMC module performs transposed convolution operation, bilinear difference processing, CBAM module attention processing and two-dimensional convolution operation on the relatively high-level features in sequence, and then fuses them element-by-element with the relatively low-level features to obtain the output of the FFMC module.

[0048] The FFMC module performs the following operations on the input features:

[0049] f out=Conv2d(CBAM(BI(TC(f high ))))+f low

[0050] Among them, f low ∈R 2H×2W×C Represents relatively low-level features, f high ∈R H×W×C Represents relatively high-level features, TC represents transposed convolution operation, BI represents bilinear difference processing, CBAM represents CBAM module attention processing operation, Conv2d represents two-dimensional convolution operation, f out ∈R 2H×2W×C Represents the features output by the FFMC module.

[0051] This example uses the DUO dataset of complex underwater scenes to train the EFE-UODNet network. The dataset is split into 80% training, 10% testing, and 10% validation sets, with input images fixed at 640×640 pixels. Mosaic data augmentation is employed during training, disabled after the 10th training cycle to prevent overfitting. Stochastic Gradient Descent (SGD) is used as the optimizer, with a fixed learning rate of 0.01 and a momentum coefficient of 0.937.

[0052] Figure 6 This is the visualization result of the ablation experiment in the present invention, where the red ellipse indicates missed detection, the yellow ellipse indicates false detection, and the black ellipse indicates repeated detection.

[0053] Table 1 Ablation experiment results

[0054] Models EGCGF Dyhead AM-FPN mAP50 mAP[50-95] IT(ms) Params(M) baseline 87.4％ 69.5％ 2.64 25.86 Model1 √ 87.7％ 70.7％ 6.33 40.99 Model2 √ √ 88.7％ 72.0％ 11.33 39.87 Model3 √ √ √ 88.9％ 73.0％ 15.30 49.05

[0055] As shown in Table 1, the EGCGF, AM-FPN, and DyHead modules all improved the performance of the baseline network. While the EGCGF and GCGF modules showed no significant change in mAP50 (mean average precision at an IoU of 0.5), mAP[50-95] (mean average precision at 0.05 intervals between 0.5 and 0.95) increased by 1.2%, demonstrating that the EGCGF module improved the model's overall detection capability for marine life. Because the EGCGF module tends to focus on and fully extract target features, the increased amount of information extracted by EGCGF reduces detection speed. Furthermore, the DyHead module achieved significant performance improvements in mean average precision (AP), increasing AP50 by 1% and mAP[50-95] by 1.3%. It effectively replaces the original detection head module without significantly increasing the number of parameters and achieves higher accuracy than the original detection head module. The AM-FPN module also performed well in terms of AP, improving mAP[50-95] by 1%.

[0056] Based on the above analysis, the designed EGCGF and AM-FPN modules are effective, and the introduced DyHead module also achieves performance improvement, which ensures the feasibility of the designed network of the application.

[0057] In order to further study the role of the proposed modules in image feature extraction, Figure 6 The performance comparison of different networks in the experiment is shown. In the experiment, multiple underwater biological images with typical complex environmental backgrounds were selected as test samples. From the detection results in the first column, it can be seen that the missed detection of sea urchins and sea cucumbers is marked with a red ellipse, and the false detection of sea urchins is marked with a yellow ellipse. As shown in the second column, after introducing the EGCGF module, the sea urchin in the top image is effectively detected, but the sea cucumber in the middle image is still not detected. At the same time, the detection error of mistaking the rock in the bottom image as a scallop is significantly improved. With the further introduction of the DyHead module, the detection accuracy of the top image in the third column is improved. However, the sea cucumber in the middle image is still missed, and although the sea cucumber in the bottom image is detected, there is a repeated detection. Finally, by introducing the AM-FPN module, the detection accuracy of the fourth column is comprehensively improved: the sea cucumber in the middle image is successfully detected, and the sea cucumber in the bottom image is also accurately identified. The experimental results show that compared with other networks, the improved network proposed in the application can achieve more accurate target prediction and higher detection score in the case of false detection and missed detection. Especially in the complex underwater environment, this method can effectively handle challenging recognition tasks such as low-contrast small targets and dense targets, and exhibits excellent detection performance.

[0058] Figure 7 is the visual detection comparison result of the application with other algorithms, wherein the red box in the figure represents the ship target detected by the model.

[0059] Table 2 comparison of experimental results

[0060]

[0061] As shown in Table 2, the table also lists the backbone structures of different networks. The effectiveness of multiple algorithms in detecting underwater organisms was verified, including the YOLO series algorithms (such as YOLOv5, YOLOv6, YOLOv7, YOLOv8n and YOLOv11n), the Faster R-CNN algorithm, the RT-DETR algorithm, the TOOD algorithm and the GCC-Net algorithm. Among them, YOLOv5m represents the medium-scale version of the algorithm, YOLOv5s represents the small-scale version of YOLO, and YOLOv5n, YOLOv8n and YOLOv11n represent the extremely small versions of the algorithm. Specifically, the method proposed in the present invention shows significant performance improvements in multiple categories such as sea cucumbers, sea urchins, scallops and starfish. For example, compared with Faster R-CNN, the present invention outperforms in mAP50, mAP[50-95] and inference time (IT) indicators. While the YOLO family of algorithms excels in terms of inference time and number of parameters, the proposed method achieves statistically significant advantages in mAP50 and mAP[50-95]. Compared to the RT-DETR and TOOD algorithms, the proposed method significantly outperforms them in both mAP50 and mAP[50-95]. While GCC-Net performs better in mAP50, its inference time is longer and it falls short of the proposed method in mAP[50-95]. Despite its faster speed, GCC-Net is less accurate than the proposed method. The proposed method achieves a nearly 2% improvement in mAP[50-95] and has a shorter inference time. In terms of mAP50, the proposed method lags slightly behind the state-of-the-art GCC-Net algorithm and is not as competitive as the nanoscale model in terms of inference time. Since real-time underwater biological detection typically requires a frame rate of at least 24 frames per second (i.e., inference time less than 42 milliseconds), the proposed algorithm achieves a relatively reasonable inference time, even faster than some other algorithms. However, the proposed method is slightly inferior to GCC-Net in terms of the number of model parameters. In summary, the significant performance improvements achieved by the proposed method highlight its significant advantages over existing technologies, demonstrating its superior accuracy and reliability in underwater object detection tasks. This method not only provides a more precise tool for the study and protection of underwater life, but also offers more reliable technical support for practical applications in marine engineering and resource management.

[0062] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. An enhanced biological image detection method for complex underwater environments, characterized in that: Abstract: In order to solve the problem of underwater biological detection error, an underwater image detection method based on the EFE-UODNet enhanced biological detection network is proposed. The method inputs the acquired underwater image into the EFE-UODNet enhanced biological detection network and outputs the underwater biological detection results. The EFE-UODNet enhanced biological detection network is improved based on the YOLOv8 network. The EGCGF module and the GCGF module are used to replace the C2F module in the backbone network for efficient multi-scale feature extraction. The AM-FPN network is used to replace the original neck network to fully utilize the context information of the underwater environment for feature fusion. The dynamic detection head is used to replace the original detection head to output the detection results through its unique multi-scale target detection mechanism.

2. The enhanced biological image detection method for complex underwater environments according to claim 1, characterized in that: The EGCGF module includes five depth-separable convolution modules and a GCGF module with large convolution kernels of different sizes. Two of the depth-separable convolution modules form a depth-separable convolution processing channel, and a total of four separable convolution processing channels and one GCGF module processing channel are constructed. The inputs are processed separately, and the five outputs are spliced ​​to obtain the output of the EGCGF module. The depth-separable convolution module consists of depth-wise convolution and point-by-point convolution. The depth-wise convolution is used to extract spatial features, and the point-by-point convolution is used to extract channel features.

3. The enhanced biological image detection method for complex underwater environments according to claim 1 or 2, characterized in that: After the GCGF module performs convolution processing on its input features, it segments the feature map to obtain a first feature map and a second feature map. The output of the second feature map after processing by N series-connected EB modules is spliced ​​with the first feature map and the second feature map, and then output after a layer of convolution.

4. The enhanced biological image detection method for complex underwater environments according to claim 3, characterized in that: The EB module is obtained by connecting the dynamic convolution and VSSBlock modules in series and then performing residual connection.

5. The enhanced biological image detection method for complex underwater environments according to claim 1, characterized in that: The EGCGF module performs the following operations on the input feature map F: Among them, Conv c1×c2 Indicates that the convolution operation with a convolution kernel size of c1×c2 is performed on the feature map, that is, Represents a two-dimensional convolution operation, i.e., convolution calculation, BN represents batch normalization, Silu represents the activation function, c1 and c2 represent the width and height of the convolution kernel respectively; N is the number of EB modules in series; Split represents the split operation; EB represents the operation of the EB module, EB(F)=VSSBlock(DynamicConv 3×3 (F))+F,DynamicConv 3×3 Indicates the dynamic convolution operation with a convolution kernel size of 3×3 on the feature map, VSSBlock indicates the operation of the VSSBlock module; GCGF(F) indicates the output of the GCGF module; DWConv c1×c2 It represents a depth-wise separable convolution operation with a convolution kernel size of c1×c2 on the feature map; Concat represents the feature map concatenation function.

6. The enhanced biological image detection method for complex underwater environments according to claim 1, characterized in that: The backbone network of the EFE-UODNet enhanced biological detection network includes two convolution modules, a first EGCGF module, a convolution module, a second EGCGF module, a convolution module, a third EGCGF module, a convolution module, a GCGF module and an SPPF module connected in sequence, wherein the first EGCGF module, the second EGCGF module, the third EGCGF module and the SPPF module output first-level features, second-level features, third-level features and fourth-level features respectively, and input them into the AM-FPN network.

7. The enhanced biological image detection method for complex underwater environments according to claim 1, characterized in that: The AM-FPN network is set with four feature processing levels, receiving the first to fourth level features output by the backbone network to capture tiny target details, small target details, medium target details and large target details respectively. The AM-FPN network includes three feature fusion modules and three feature extraction modules. The first level features are input into the first feature fusion module, the second level features are input into the second feature fusion module, the third level features are input into the third feature fusion module, and the fourth level features are input into the third feature fusion module and the third feature extraction module at the same time. The output of the third feature fusion module is input into the second feature extraction module, and the output of the first and second feature fusion modules are input into the first feature extraction module. The outputs of the first to third feature extraction modules and the input of the first feature extraction module constitute the four outputs of the AM-FPN network, which are respectively input into the corresponding four dynamic detection heads.

8. The enhanced biological image detection method for complex underwater environments according to claim 7, characterized in that: The feature fusion module is composed of a cascade of an FFMC module and a C2F module, and the feature extraction module is composed of a cascade of a Conv layer, a Concat layer and a C2F module.

9. The enhanced biological image detection method for complex underwater environments according to claim 8, characterized in that: The FFMC module includes two inputs: relatively low-level features and relatively high-level features. The FFMC module performs transposed convolution operation, bilinear interpolation processing, CBAM module attention processing and two-dimensional convolution operation on the relatively high-level features in sequence, and then fuses them element-by-element with the relatively low-level features to obtain the output of the FFMC module.

10. The enhanced biological image detection method for complex underwater environments according to claim 9, characterized in that: The FFMC module performs the following operations on the input features: f out =Conv2d(CBAM(BI(TC(f high ))))+f low Among them, f low ∈R 2H×2W×C Represents relatively low-level features, f high ∈R H×W×C Represents relatively high-level features, TC represents transposed convolution operation, BI represents bilinear difference processing, CBAM represents CBAM module attention processing operation, Conv2d represents two-dimensional convolution operation, f out ∈R 2H×2W×C Represents the features output by the FFMC module.

Images