Underwater fishing net detection method and device based on YOLOv11

Abstract

The invention provides a YOLOv11-based underwater fishing net detection method and device, and the method comprises the steps: carrying out the feature extraction of an inputted to-be-detected underwater image through a trunk feature extraction network in an improved YOLOv11 detection model, obtaining an initial feature map, and employing a MobileNetV2 network as the trunk feature extraction network; multi-scale feature fusion is carried out on the initial feature map through a neck network in the detection model, a fused feature map is obtained, and the neck network adopts a Slim-Neck structure; performing fishing net detection on the fusion feature map through a detection head network in the detection model to obtain a detection result of a fishing net target; and the detection head network adopts a GSConv module, a GSBottleneck bottleneck layer and a VoV-GSCSP module, so that the reasoning speed and the detection efficiency of the model are improved, the integration capability of the model for different-scale fishing net features is enhanced, the robustness of the model in a complex underwater scene is improved, and relatively high detection precision is maintained on the premise of improving the efficiency.

Description

Technical Field

[0001] This invention relates to the field of underwater image detection technology, specifically to an underwater fishing net detection method and device based on YOLOv11. Background Technology

[0002] Unmanned Underwater Vehicles (UUVs), as highly intelligent operational platforms, play an increasingly important role in marine exploration, environmental monitoring, and underwater facility inspection and maintenance. Ensuring the safe navigation of UUVs is a prerequisite for them to perform various tasks. However, in the complex real-world marine environment, especially in nearshore and aquaculture areas, there are numerous fishing nets abandoned or damaged by fishing or aquaculture activities. These nets are often suspended or semi-suspended, with weak visual characteristics and irregular shapes, making them highly susceptible to entanglement in the UUV's propellers or sensors, leading to equipment damage, mission failure, and even safety accidents. Therefore, achieving real-time and accurate detection of fishing nets in the underwater environment is of crucial practical significance for the autonomous obstacle avoidance and safe navigation of UUVs.

[0003] In related technologies, underwater fishing net detection methods mostly employ deep learning-based target detection algorithms, particularly models such as YOLO (You Only Look Once), SSD (Single Shot MultiBox Detector), and Faster R-CNN. However, directly applying these models to underwater fishing net detection still has significant shortcomings: On the one hand, mainstream deep learning models typically have high computational complexity and a large number of parameters, requiring powerful computing platforms, which contradicts the limited onboard computing resources of embedded devices such as UUVs; on the other hand, these models are not designed for underwater optical imaging characteristics and lack robustness to underwater image degradation problems (such as fogging and color cast), resulting in decreased detection performance in complex underwater environments. In particular, the detection accuracy and recall rate for fine, semi-transparent, and low-contrast targets such as fishing nets are difficult to meet practical requirements, reducing the detection efficiency and robustness of underwater fishing net target detection. Summary of the Invention

[0004] In view of this, it is necessary to provide an underwater fishing net detection method and device based on YOLOv11 to solve the technical problem of low detection efficiency and robustness caused by using the existing YOLOv11 model for underwater fishing net target detection.

[0005] To address the aforementioned technical problems, in a first aspect, the present invention provides an underwater fishing net detection method based on YOLOv11, comprising: The improved YOLOv11 detection model uses a backbone feature extraction network to extract features from the input underwater image to be detected, resulting in an initial feature map. The backbone feature extraction network is replaced by a MobileNetV2 network instead of the original YOLOv11 backbone feature extraction network. The initial feature map is fused using the neck network in the improved YOLOv11 detection model to obtain a fused feature map. The neck network uses a Slim-Neck structure to replace the original YOLOv11 neck network. The detection head network in the improved YOLOv11 detection model is used to detect fishing nets in the fused feature map to obtain the detection result of the fishing net target; wherein, the detection head network uses the GSConv module, the GSBottleneck bottleneck layer and the VoV-GSCSP module to replace the original YOLOv11 detection head network.

[0006] In one possible implementation, prior to feature extraction of the input underwater image via the backbone feature extraction network in the improved YOLOv11 detection model, the method further includes: Raw underwater images are acquired using an underwater robot equipped with a monocular camera and an auxiliary light source. An improved dark channel prior enhancement algorithm is used to perform color enhancement processing on the original underwater image to generate the underwater image to be detected.

[0007] In one possible implementation, the step of using an improved dark channel prior enhancement algorithm to perform dehazing and color enhancement processing on the original underwater image to generate the underwater image to be detected includes: The input raw underwater image is subjected to channel-adaptive dark channel estimation based on the minimum values ​​of the blue and green channels to obtain a preliminary transmittance map; Bright area extraction and background light estimation are performed on the preliminary transmittance map to obtain background light values ​​including red channel ratio correction; The preliminary transmittance map is subjected to guided filtering based on Sobel edge weights to obtain an optimized transmittance map; Based on the optimized transmittance map, relative depth calculation and adaptive gain compensation processing of the red channel are performed to obtain the compensated image, wherein the gain value of the adaptive gain compensation of the red channel is less than the adaptive upper limit related to transmittance. The compensated image is then subjected to white balance adjustment, CLAHE-based local contrast enhancement, and texture sharpening to obtain the underwater image to be detected.

[0008] In one possible implementation, the MobileNetV2 network is an inverse residual structure, including a 1×1 convolutional expansion layer, a depthwise separable convolutional layer, and a 1×1 convolutional reduction layer; the feature extraction of the input underwater image through the backbone feature extraction network in the improved YOLOv11 detection model to obtain an initial feature map includes: The underwater image to be detected is subjected to channel expansion processing through the 1×1 convolutional expansion layer to obtain a high-dimensional feature map. The high-dimensional feature map is processed by the depthwise separable convolutional layer to extract spatial features, thus obtaining a spatial feature map. The initial feature map is obtained by performing channel compression on the spatial feature map through the 1×1 convolutional reduction layer.

[0009] In one possible implementation, the fishnet detection of the fused feature map using the detection head network in the improved YOLOv11 detection model includes: The GSConv module is used to perform feature transformation on the fused feature map to obtain a first feature map; The first feature map is input into the GSBottleneck bottleneck layer for feature compression and cross-layer information fusion to obtain the second feature map; The second feature map is input into the VoV-GSCSP module for multi-scale feature enhancement and integration to obtain an enhanced multi-scale feature map, which is then used to generate the detection result of the fishing net target.

[0010] In one possible implementation, the step of using the GSConv module to perform feature transformation on the fused feature map to obtain a first feature map includes: The fused feature map is divided into first path features and second path features; Perform a standard convolution operation on the first path features and a depthwise separable convolution operation on the second path features; The first path feature after the standard convolution operation and the second path feature after the depthwise separable convolution operation are concatenated along the channel dimension to obtain the first concatenated feature. Perform a channel shuffling operation on the first spliced ​​feature to generate the first feature map.

[0011] In one possible implementation, the GSBottleneck bottleneck layer includes a preset GSConv module; the step of inputting the first feature map into the GSBottleneck bottleneck layer for feature compression and cross-layer information fusion to obtain a second feature map includes: The first feature map is input into the preset GSConv module in the GSBottleneck bottleneck layer for feature compression processing to obtain the basic transformation features. The basic transformation features are subjected to residual connection and channel integration processing to obtain the second feature map.

[0012] In one possible implementation, the detection head network is built on the SSDLite framework, in which the convolutional networks include depthwise separable convolutional networks.

[0013] In one possible implementation, the step of inputting the second feature map into the VoV-GSCSP module for multi-scale feature enhancement and integration to obtain an enhanced multi-scale feature map includes: The second feature map is divided into main path features and auxiliary path features; Perform a standard convolution operation on the main path features; The auxiliary path features are subjected to standard convolution operations and lightweight bottleneck layer processing; The main path features after standard convolution operation and the auxiliary path features after lightweight bottleneck layer processing are concatenated in the channel dimension to obtain the second concatenated feature. Perform a 3×3 convolution operation on the second spliced ​​feature to obtain an enhanced multi-scale feature map.

[0014] Secondly, the present invention also provides an underwater fishing net detection device based on YOLOv11, comprising: The extraction unit is used to extract features from the input underwater image to be detected through the backbone feature extraction network in the improved YOLOv11 detection model to obtain an initial feature map. The backbone feature extraction network adopts MobileNetV2 network instead of the original YOLOv11 backbone feature extraction network. The fusion unit is used to perform multi-scale feature fusion on the initial feature map through the neck network in the improved YOLOv11 detection model to obtain a fused feature map. The neck network adopts a Slim-Neck structure to replace the original YOLOv11 neck network. The detection unit is used to perform fishing net detection on the fused feature map through the detection head network in the improved YOLOv11 detection model to obtain the detection result of the fishing net target; wherein, the detection head network adopts the GSConv module, the GSBottleneck bottleneck layer and the VoV-GSCSP module to replace the original YOLOv11 detection head network.

[0015] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the YOLOv11-based underwater fishing net detection method described in any of the above implementations.

[0016] Fourthly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instructions, which, when executed by a processor, can implement the steps of the underwater fishing net detection method based on YOLOv11 described in any of the above implementations.

[0017] The beneficial effects of this invention are: The underwater fishing net detection method based on YOLOv11 provided by this invention extracts features from the input underwater image to be detected through an improved backbone feature extraction network in the YOLOv11 detection model to obtain an initial feature map. The backbone feature extraction network is replaced by a MobileNetV2 network instead of the original YOLOv11 backbone feature extraction network. Multi-scale feature fusion is performed on the initial feature map through a neck network in the improved YOLOv11 detection model to obtain a fused feature map. The neck network is replaced by a Slim-Neck structure instead of the original YOLOv11 neck network, fully utilizing the combined design features of the inverse residual and linear bottleneck layer of the MobileNetV2 network. This achieves flexible feature dimension transformation and efficient extraction while ensuring the stability of the feature information flow, which is beneficial to improving the robustness of fishing net detection. The detection head network in the advanced YOLOv11 detection model performs fishing net detection on the fused feature map to obtain the detection result of the fishing net target. By optimizing the path and reducing redundant calculations, an efficient information flow is established between feature maps of multiple scales, improving the feature transfer efficiency. It can reduce the computational complexity (GFlops) while ensuring detection accuracy, thereby improving the overall computational efficiency. In particular, the detection head network uses the GSConv module, GSBottleneck bottleneck layer and VoV-GSCSP module to replace the original YOLOv11 detection head network, which significantly reduces the computational complexity of the entire detection head network, thereby improving the model's inference speed and detection efficiency. It also strengthens the model's ability to integrate fishing net features of different scales, improves the model's robustness in complex underwater scenes, and maintains high detection accuracy while improving efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic flowchart of an embodiment of the underwater fishing net detection method based on YOLOv11 provided by the present invention; Figure 2 This is a schematic diagram of the process for generating an underwater image to be detected, provided by the present invention. Figure 3 For the present invention Figure 2 A schematic diagram of an embodiment of S202; Figure 4 A comparison image of the original underwater image and the underwater image to be detected provided by this invention; Figure 5 Provided by the present invention Figure 1 A schematic flowchart of an embodiment of S103 Figure 6 A schematic diagram illustrating the principle of the underwater fishing net detection method based on YOLOv11 provided by this invention; Figure 7 A schematic diagram of the YOLOv11 network structure in the improved YOLOv11 detection model provided by this invention; Figure 8 This is a schematic diagram of the GSBottleneck module structure provided by the present invention; Figure 9 A schematic diagram of the VoV-GSCSP module structure provided by the present invention; Figure 10 This is a schematic diagram of the underwater fishing net detection device based on YOLOv11 provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] In the description of the embodiments of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0022] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] This invention provides an underwater fishing net detection method and device based on YOLOv11, which will be described below.

[0025] The execution entity of the underwater fishing net detection method based on YOLOv11 in this application embodiment can be the underwater fishing net detection device based on YOLOv11 provided in this application embodiment, or different types of electronic devices such as server equipment, physical host, or user equipment (UE) that integrate the underwater fishing net detection device based on YOLOv11. The underwater fishing net detection device based on YOLOv11 can be implemented in hardware or software. The UE can be a terminal device such as a smartphone, tablet computer, laptop computer, handheld computer, desktop computer, or personal digital assistant (PDA).

[0026] Figure 1 This is a schematic flowchart of an embodiment of the underwater fishing net detection method based on YOLOv11 provided by the present invention, as shown below. Figure 1 As shown, the underwater fishing net detection method based on YOLOv11 includes: S101. The input underwater image to be detected is used to extract features through the backbone feature extraction network in the improved YOLOv11 detection model to obtain an initial feature map. The backbone feature extraction network is replaced by the MobileNetV2 network instead of the original YOLOv11 backbone feature extraction network.

[0027] The YOLOv11 (You Only Look Once version 11) detection model is a single-stage deep learning model for object detection, which achieves a balance between speed and accuracy. The original YOLOv11 backbone network is replaced with MobileNetV2, employing an inverse residual structure and a linear bottleneck layer to reduce computational cost.

[0028] The backbone network, also known as the core network, is the front-end of a deep neural network. It is used to extract and condense basic visual features (such as edges and textures) and advanced semantic features from the input image layer by layer.

[0029] MobileNetV2 is a lightweight convolutional neural network architecture whose core consists of an inverted residual structure and a linear bottleneck layer, which can maintain high accuracy while significantly reducing computation and the number of parameters.

[0030] The inventors discovered that the original YOLOv11 detection model typically uses a conventional convolutional neural network (such as the CSPDarknet architecture) with a high computational load as the backbone network. Although it can effectively extract features, the number of parameters and computational complexity are high, which is not conducive to deployment on resource-constrained embedded platforms.

[0031] Specifically, this embodiment uses the MobileNetV2 network as the backbone feature extraction network. That is, the lightweight target detection network MobileNetV2 is used to replace the backbone feature extraction network in the original YOLOv11. Multi-level feature extraction and downsampling are performed on the input underwater image to be detected, and the final output is an initial feature map containing high-level semantic information. Since the MobileNetV2 network has a linear bottleneck layer, the destruction of feature information in the low-dimensional space is avoided by removing the traditional non-linear activation function, and the linear mapping relationship of the feature representation can be maintained, thereby improving the feature preservation ability and detection performance of the network. This embodiment makes full use of the combined design features of the inverse residual and the linear bottleneck layer of the MobileNetV2 network. While achieving flexible transformation and efficient extraction of feature dimensions, it ensures the stability of the feature information flow, which is conducive to improving the robustness of fishing net detection.

[0032] S102. Multi-scale feature fusion is performed on the initial feature map through the neck network in the improved YOLOv11 detection model to obtain a fused feature map. The neck network adopts a Slim-Neck structure to replace the original YOLOv11 neck network.

[0033] The neck network, located between the backbone network and the detection head network, is used to fuse and enhance feature maps of different scales extracted by the backbone network to construct a feature pyramid rich in multi-scale information. In this embodiment, in the improved YOLOv11 detection model, the neck network is used to perform multi-scale feature fusion on the initial feature maps extracted by the backbone network.

[0034] Specifically, a Slim-Neck structure is used to replace the feature fusion layer (such as FPN or PANet) of the original model. This structure establishes an efficient information flow between feature maps of multiple scales by optimizing the path and reducing redundant computation, thereby improving feature transfer efficiency. It can reduce computational complexity (GFlops) while ensuring detection accuracy, thus improving overall computational efficiency.

[0035] In one specific implementation, Slim-Neck can be a depthwise separable convolution or attention mechanism to fuse feature maps at different scales while reducing the number of parameters. Slim-Neck receives feature maps from three different layers of the backbone network (e.g., features downsampled by 8x, 16x, and 32x). First, it applies lightweight convolutional modules (such as depthwise separable convolution) to each scale feature map for channel adjustment and feature extraction. Then, through a bidirectional path from top to bottom and bottom to top, it progressively fuses and transfers deep semantic features with shallow detail features, ultimately outputting a set of fused feature maps with rich semantic information and high spatial resolution. In the task of detecting fine textures and complex backgrounds of underwater fishing nets, it can significantly reduce the computational burden of the feature fusion process while ensuring that fishing net targets at different scales (such as large meshes in the foreground and blurred net lines in the distance) receive sufficient feature representation. Experiments have shown that the Slim-Neck structure reduces the overall computational complexity (GFLOPs) of the model by 38% while sacrificing no more than 1% in detection accuracy (mAP50). This directly leads to a significant improvement in the model's inference speed and lays a key technological foundation for deploying the model on underwater embedded devices with limited computing resources and meeting real-time detection requirements.

[0036] S103. The fused feature map is used to detect fishing nets through the detection head network in the improved YOLOv11 detection model to obtain the detection result of the fishing net target; wherein, the detection head network uses the GSConv module, the GSBottleneck bottleneck layer and the VoV-GSCSP module to replace the original YOLOv11 detection head network.

[0037] The detection head network is the last part of the target detection network. It receives the fused features output by the neck network and performs specific classification and localization tasks.

[0038] The Group Shuffle Convolution (GSConv) module is a lightweight hybrid convolution module that splits the input feature map into two parallel paths: one path uses standard convolution to extract main features, and the other path uses depthwise separable convolution to extract detailed features. Finally, the two outputs are concatenated along the channel dimension and channel shuffled to achieve efficient information exchange.

[0039] The GSBottleneck Layer is a composite module with a bottleneck structure (compression-processing-expansion) with the GSConv module as its core building unit. It is used to achieve efficient information fusion between feature maps of different levels while maintaining lightweight characteristics.

[0040] The VoV-GhostNet Cross Stage Partial Connection Module (VoV-GSCSP module) based on the VoV structure is a complex feature fusion module that contains multiple parallel paths (such as main paths and auxiliary paths). These paths may integrate conventional convolutional layers, bottleneck layers, and other structures. Finally, they are concatenated and convolved to enhance the network's multi-scale expressive capabilities.

[0041] Specifically, the GSConv module is first used to perform preliminary feature transformation on the fused feature map. This module divides the input feature map into two parallel paths: one path performs standard convolution to capture the main outline, and the other path performs depthwise separable convolution to extract local details. Channel concatenation and shuffling are then performed to achieve efficient feature reuse and information exchange, outputting the first feature map. Subsequently, the first feature map is input into the GSBottleneck bottleneck layer. This layer uses the GSConv operation as its core processing unit to compress and nonlinearly transform the input features, and combines residual connection mechanisms to fuse contextual information from different network depths, outputting a second feature map with denser information and richer semantics. Finally, the second feature map is input into the VoVGSCSP module for multi-scale feature enhancement and integration. This module divides the input features into a main path and an auxiliary path. The main path retains the main features through standard convolution, while the auxiliary path extracts supplementary detailed features through cascaded processing of standard convolution and a lightweight bottleneck layer. The outputs of the two paths are then concatenated along the channel dimension and fused using a 3×3 convolution to output an enhanced multi-scale feature map. This feature map is directly used for subsequent bounding box regression and classification to generate the final fishing net target detection result. This significantly reduces the computational complexity of the entire detection head network, thereby improving the model's inference speed and detection efficiency. It also enhances the model's ability to integrate features from fishing nets of different scales, improving its robustness in complex underwater scenarios and meeting the stringent real-time detection requirements of underwater embedded devices. Simultaneously, the lightweight design reduces computational load while effectively preserving, fusing, and enhancing feature information, ensuring the model's ability to capture the fine, low-contrast, and morphologically variable characteristics of underwater fishing nets, thus maintaining high detection accuracy while improving efficiency.

[0042] In one specific implementation, an underwater image that has been dehazed and enhanced is input. The fused feature map obtained after processing by the backbone network and the neck network is then processed sequentially through the three detection modules mentioned above. Finally, the precise bounding box coordinates and confidence scores of all fishing net targets in the image can be output.

[0043] In summary, the underwater fishing net detection method based on YOLOv11 provided in this embodiment of the invention extracts features from the input underwater image to be detected through the backbone feature extraction network in the improved YOLOv11 detection model to obtain an initial feature map. The backbone feature extraction network is replaced by a MobileNetV2 network instead of the original YOLOv11 backbone feature extraction network. The initial feature map is then fused using a multi-scale feature fusion method through the neck network in the improved YOLOv11 detection model to obtain a fused feature map. The neck network is replaced by a Slim-Neck structure instead of the original YOLOv11 neck network. This fully utilizes the combined design features of the inverse residual and linear bottleneck layer of the MobileNetV2 network, achieving flexible feature dimension transformation and efficient extraction while ensuring the stability of the feature information flow, thus improving the robustness of fishing net detection. The improved YOLOv11 detection model uses a modified detection head network to detect fishing nets from fused feature maps. By optimizing paths and reducing redundant computations, an efficient information flow is established between feature maps of multiple scales, improving feature transfer efficiency. This reduces computational complexity (GFlops) while ensuring detection accuracy, thus improving overall computational efficiency. Specifically, the detection head network replaces the original YOLOv11 detection head network with the GSConv module, the GSBottleneck bottleneck layer, and the VoV-GSCSP module, significantly reducing the computational complexity of the entire detection head network. This improves the model's inference speed and detection efficiency, strengthens its ability to integrate features of fishing nets at different scales, and enhances its robustness in complex underwater scenarios, maintaining high detection accuracy while improving efficiency.

[0044] In some embodiments of the present invention, such as Figure 2 As shown, before step S101, the procedure further includes: S201. Acquire raw underwater images using an underwater robot equipped with a monocular camera and an auxiliary light source; S202. An improved dark channel prior enhancement algorithm is used to perform color enhancement processing on the original underwater image to generate the underwater image to be detected.

[0045] One approach is to equip a spectral monocular camera on a Remotely Operated Vehicle (ROV) as the core sensor. The auxiliary light source can be a linear laser installed on the ROV, which can introduce active illumination in low-light environments. This allows for the design of acquisition schemes from three aspects: lighting conditions, shooting distance, and shooting angle, in order to improve the generalization of the acquired raw underwater images.

[0046] Specifically, an improved DCP (Dark Channel Prior) algorithm tailored to the underwater environment can be used to enhance and preprocess the original underwater image, thereby removing fog from the original underwater image and improving the quality of the underwater image to be detected.

[0047] In some embodiments of the present invention, such as Figure 3 As shown, step S202 includes: S301. Perform channel adaptive dark channel estimation processing on the input raw underwater image based on the minimum values ​​of the blue and green channels to obtain a preliminary transmittance map; S302. Perform bright area extraction and background light estimation on the preliminary transmittance map to obtain a background light value including red channel ratio correction. S303. Perform guided filtering based on Sobel edge weights on the preliminary transmittance map to obtain an optimized transmittance map. S304. Based on the optimized transmittance map, perform relative depth calculation and red channel adaptive gain compensation processing to obtain the compensated image, wherein the gain value of the red channel adaptive gain compensation is less than the adaptive upper limit related to transmittance. S305. The compensated image is sequentially subjected to white balance adjustment, local contrast enhancement based on CLAHE, and texture sharpening to obtain the underwater image to be detected.

[0048] Among these features, adaptive dark channel estimation uses only the blue and green channels to estimate transmittance, avoiding the attenuation effect of the red channel. Edge-guided filtering is used in conjunction with Sobel edge weights to enhance the fishing net structure. Background light estimation is biased towards the blue and green components, with the red channel compensated proportionally.

[0049] Adaptive parameter adjustment and CLAHE enhancement are employed, which involves adjusting parameters based on image statistical features to enhance local contrast.

[0050] Red channel adaptive gain compensation processing refers to balancing colors using LAB mode based on depth adaptive gain.

[0051] Specifically, in this embodiment, image detail enhancement and color balance are achieved by combining channel adaptive modeling, edge-guided filtering, red light compensation, white balance adjustment, and adaptive parameter optimization.

[0052] In one specific implementation, channel-adaptive dark channel estimation is proposed. Traditional DCP uses the minimum value of the RGB three channels as the dark channel, but this is not suitable for underwater environments where red light attenuates too quickly. This embodiment proposes a new estimation method:

[0053] in, For pixel position x The estimated blue-green dark channel value represents the local region Ω ( x The lowest level of intensity in the inner blue and green channels is used to replace the level that might be underestimated due to red light attenuation in traditional algorithms. RGB The minimum value of the three channels is used to more accurately characterize the transmission characteristics of underwater scenes; Ω( x (in pixels) x Minimum operations are performed on a local image region centered on the image to capture local light attenuation characteristics. G '( y ), B '( y ): In position y The preprocessed green and blue channel intensity values ​​are used in the formula; the transmittance is estimated using only the blue and green channels, which is more consistent with the characteristics of underwater images.

[0054] Background light estimation uses bright areas of the image with minimum blue-green channel values ​​to extract candidate background light pixels. Background light A is biased towards the blue-green component, and the red component is adjusted proportionally to maintain balance.

[0055]

[0056] in, , is the ambient light ratio correction coefficient for the red channel. Since red light decays the fastest underwater, the directly estimated red background light value is extremely low, which will cause the restored image to be severely blue. γ is introduced to proportionally increase the red component (for example, to restore it based on half of the average blue and green values) in order to balance the color and avoid producing unnatural excessive color cast. A The estimated background light vector is a three-dimensional vector, corresponding to the background light intensity of the red (R), green (G), and blue (B) channels, respectively; For each candidate pixel extracted from the bright area of ​​the dark channel image, calculate the average intensity of its green and blue channels.

[0057] Edge-guided filtering is used in fishing net images to smooth the transmittance, where a finer mesh structure is required.

[0058] Seek Then, a weighted average is calculated, from which the following can be obtained:

[0059] Where I is the brightness of the guide image; To in a local window The energy function within the function is used to solve for the optimal linear transformation coefficients by minimizing this function. and ; The output transmittance after filtering; To guide the image (usually the luminance channel of the original image) at position i The intensity value at that location; For the preliminary transmittance map at location i The value at; For regularization parameters; Used to prevent coefficient If the value is too large, it will act as a smoothing constraint, preventing the filtering result from overfitting noise or details and maintaining overall smoothness.

[0060] We enhance “line attention” by introducing Sobel edge weights E(x).

[0061]

[0062] Optimized transmittance after edge enhancement; This is to guide the initial smooth transmission after filtering; This refers to the edge intensity weight map calculated using edge operators such as Sobel, located at the image edges. The value is relatively large, and the value is relatively small in flat areas; This is the edge enhancement intensity coefficient, used to control the enhancement of the transmittance map by edge information; The higher the value, the more the transmittance of the edge area is enhanced, thus better preserving and highlighting the contrast of detailed structures such as fishing net lines in the final dehazed image.

[0063] Red channel compensation, using relative depth z= ln(tG) defines the red channel gain:

[0064] Set adaptive upper limits for brightness and transmittance:

[0065] For in position x The gain value used for adaptive gain compensation of the red channel; Based on green channel transmittance The calculated relative depth estimate ( ln(tG)), The larger the value, the deeper the estimated depth of the region and the more severe the red light attenuation. This is the gain adjustment coefficient. It is a proportionality constant, where Δβ represents the difference in attenuation coefficients between red and green light. The product of the two, κ·Δβ, jointly determines the gain. With depth The rate of exponential growth is used to compensate for the exponential decay characteristics of red light, guided by a physical model. This is the transmittance estimate for the green channel. The green channel was chosen because it exhibits moderate attenuation underwater, is less prone to excessive scattering than blue light, and is more reliable than red light. It is commonly used to represent the overall depth of a scene. This constitutes the upper limit of the gain, representing the gain at close range of the object ( When the value is close to 1, the maximum allowable gain is relatively small; in distant views or dense fog areas ( When the gain is close to 0, the maximum allowable gain increases (but does not exceed 1.8). This dynamic upper limit effectively prevents overcompensation in already bright areas or due to estimation errors, thereby avoiding red oversaturation and color distortion.

[0066] White balance adjustment: Use V2 mode (LAB red reduction + G / B priority) to balance the colors.

[0067] Adaptive parameter adjustment: The proportion of channels in each fishing net image is different, so the parameters are adjusted according to the statistical characteristics of the image.

[0068] The white balance based on the grayscale world assumption, along with the CLAHE local contrast enhancement module, brings the average value of the three channels close to gray, achieving color averaging. CLAHE is applied only to dark images to enhance local contrast and suppress haze effects, while Unsharp Mask is introduced to slightly enhance texture details. Figure 4 The image shown is a comparison between the original underwater image and the underwater image to be detected. The left image is the original underwater image, and the right image is the underwater image to be detected.

[0069] In some embodiments of the present invention, the MobileNetV2 network is an inverse residual structure, including a 1×1 convolutional expansion layer, a depthwise separable convolutional layer, and a 1×1 convolutional reduction layer; step S101 includes: performing channel expansion processing on the underwater image to be detected through the 1×1 convolutional expansion layer to obtain a high-dimensional feature map; performing spatial feature extraction processing on the high-dimensional feature map through the depthwise separable convolutional layer to obtain a spatial feature map; and performing channel compression processing on the spatial feature map through the 1×1 convolutional reduction layer to obtain the initial feature map.

[0070] The MobileNetV2 network adopts an inverted residual structure, meaning that the MobileNetV2 network uses an inverted residual structure and a linear bottleneck layer as its core building blocks.

[0071] Specifically, the inverse residual structure first expands the number of feature channels using 1×1 convolutions, then uses depthwise separable convolutions for spatial feature extraction, and finally compresses the number of channels using another 1×1 convolution. The linear bottleneck layer refers to using a linear activation function after the final 1×1 convolution compression to avoid feature information being destroyed by nonlinear functions in low-dimensional space. For example, for an input feature map, this structure can first expand its number of channels from 24 to 144, process it with depthwise separable convolutions, and then compress it back to 24 channels. This significantly reduces computation while maintaining feature expressiveness, achieving efficient feature extraction and parameter compression, and significantly reducing computational costs.

[0072] Understandably, replacing the backbone network with the MobileNetV2 network offers several advantages. First, the use of depthwise separable convolution fundamentally breaks down standard convolution into two steps: depthwise convolution and pointwise convolution, significantly reducing the number of model parameters and floating-point computation. Second, the combined design of inverse residuals and linear bottlenecks ensures the stability of the feature information flow while enabling flexible transformation and efficient extraction of feature dimensions.

[0073] In some embodiments of the invention, such as Figure 5 As shown, step S103 includes: S401. The GSConv module is used to perform feature transformation on the fused feature map to obtain a first feature map; S402. Input the first feature map into the GSBottleneck bottleneck layer for feature compression and cross-layer information fusion to obtain the second feature map; S403. Input the second feature map into the VoV-GSCSP module for multi-scale feature enhancement and integration to obtain an enhanced multi-scale feature map, which is used to generate the detection result of the fishing net target.

[0074] Specifically, the GSConv module is used to perform feature transformation on the fused feature map. Compared with the traditional convolutional structure, the GSConv module significantly reduces the number of parameters and computational complexity while maintaining the richness of output features, achieving efficient feature expression in a lightweight network, and thus realizing efficient basic feature transformation. The first feature map is input into the GSBottleneck bottleneck layer, which uses the GSConv module as its core structural unit to construct a "compression-processing-expansion" bottleneck structure containing residual connections, realizing cross-layer deep feature fusion. Subsequently, the compressed features are fused with the original input or intermediate state of the module through residual connections to achieve cross-layer information transfer and feature reconstruction, and finally output the second feature map. The second feature map is input into the VoV-GSCSP module for multi-scale feature enhancement and integration, which enhances the robustness of the model to fishing net targets of different scales (such as large meshes in the foreground and small meshes in the background). Its output feature map contains more comprehensive and discriminative multi-scale information.

[0075] In some embodiments of the present invention, step S401 includes: dividing the fused feature map into a first path feature and a second path feature; performing a standard convolution operation on the first path feature and a depthwise separable convolution operation on the second path feature; concatenating the first path feature after the standard convolution operation and the second path feature after the depthwise separable convolution operation in the channel dimension to obtain a first concatenated feature; and performing a channel shuffling operation on the first concatenated feature to generate the first feature map.

[0076] Specifically, the GSConv module's partitioning ratio, for example, by dividing the fused feature map into first-path features and second-path features based on the average number of channels or a specific ratio, is considered. The GSConv module employs an efficient hybrid convolutional structure, which works by dividing the input feature map into two parts along the channel dimension, forming two parallel processing paths. The first path extracts global main features through a standard convolutional layer; the second path extracts local detail features through a depthwise separable convolutional layer. This operation significantly reduces parameters and computational cost. The feature maps output from the two paths are concatenated along the channel dimension, and a channel shuffle operation is performed to promote interaction and fusion of information from different paths. Thus, the output features of the two paths are concatenated (Concat) along the channel dimension and rearranged through a channel shuffle operation to achieve cross-channel information interaction and fusion.

[0077] In some embodiments of the present invention, the GSBottleneck bottleneck layer includes a preset GSConv module; step S402 includes: inputting the first feature map into the preset GSConv module in the GSBottleneck bottleneck layer for feature compression processing to obtain basic transformation features; performing residual connection and channel integration processing on the basic transformation features to obtain the second feature map.

[0078] Specifically, the first input feature map is nonlinearly transformed and compressed using the pre-defined GSConv module built into the GSBottleneck bottleneck layer. Then, the compressed features are fused with the original input or intermediate state of the module through residual connections, achieving cross-layer information transfer and feature reconstruction, ultimately outputting the second feature map. This approach not only inherits the lightweight advantages of the GSConv module but also enhances gradient flow and feature reuse through residual structures, effectively fusing shallow detail information with deep semantic information. This significantly improves the feature extraction capability and model training stability for targets with complex and fine textures, such as underwater fishing nets.

[0079] Understandably, this embodiment further introduces a GSBottleneck bottleneck layer, which uses GSConv as the core structural unit, to achieve information fusion between feature maps at different stages. This structure enhances the semantic correlation between shallow and deep features while maintaining lightweight characteristics, thereby improving the model's feature extraction capability and real-time processing performance when detecting finely textured targets (such as underwater fishing nets).

[0080] In some embodiments of the present invention, the detection head network is built based on the SSDLite framework, in which the convolutional network includes a depthwise separable convolutional network.

[0081] Specifically, the detection head network adopts the SSDLite framework, which significantly reduces model parameters and computational cost by using depthwise separable convolutions instead of standard convolution operations. The SSDLite framework is based on a single-shot detection mechanism, which improves detection speed while maintaining high detection accuracy, making it suitable for embedded devices with limited computing resources and real-time detection scenarios.

[0082] In some embodiments of the present invention, step S403 includes: dividing the second feature map into main path features and auxiliary path features; performing a standard convolution operation on the main path features; performing a standard convolution operation and lightweight bottleneck layer processing on the auxiliary path features; concatenating the main path features after the standard convolution operation and the auxiliary path features after the lightweight bottleneck layer processing in the channel dimension to obtain a second concatenated feature; and performing a 3×3 convolution operation on the second concatenated feature to obtain an enhanced multi-scale feature map.

[0083] Specifically, the input second feature map is divided into main path features and auxiliary path features along the channel dimension. The main path quickly extracts and retains core features through a standard convolutional layer, while the auxiliary path sequentially passes through a standard convolutional layer and a lightweight bottleneck layer containing residual connections for deeper feature transformation and information compression. The features processed by the two paths are then re-concatenated along the channel dimension and fused and adjusted through a 3×3 convolutional layer, finally outputting an enhanced multi-scale feature map. This structural approach enhances the model's robustness to fishing net targets of different scales through divide-and-conquer differentiated processing and efficient integration of the feature maps. The output feature map contains more comprehensive and discriminative multi-scale information, directly providing key input for subsequent classification and localization layers to generate accurate fishing net target detection results.

[0084] Understandably, this embodiment divides the input feature map into two parallel paths: the main path extracts the main features through conventional convolution; the auxiliary path sequentially passes through conventional convolution and a lightweight bottleneck layer (containing residual connection structures) to achieve feature compression and reconstruction; the output results of the two paths are concatenated in the channel dimension, and the output feature map is generated through 3×3 enhanced convolution, which improves the feature fusion effect and the network's multi-scale expression capability, thereby further improving the detection accuracy and model stability.

[0085] In one specific implementation, such as Figure 6 The diagram shows the principle block of the underwater fishing net detection method based on YOLOv11. Specifically, it involves a novel dehazing method based on dark channel priors, followed by annotation, and then training the YOLOv11 model using a dataset. Table 1 shows a performance comparison between YOLOv11 and the improved YOLOv11 detection model in this embodiment. Experimental results show that the overall computational complexity (measured in GFLOPs) of the model in this application decreases from 6.3 to 3.9, a reduction of approximately 38%, while the average precision (mAP) loss on the standard underwater dataset is controlled within 1%. The improved YOLOv11 detection model in this embodiment maintains high detection accuracy under different lighting conditions, distances, and angles, demonstrating good robustness and practical application potential.

[0086] Table 1. Performance Comparison Data of YOLOv11 and the Improved YOLOv11 Detection Model in this Embodiment

[0087] In another specific implementation, such as Figure 7 The image shows a schematic diagram of the YOLOv11 network structure in the improved YOLOv11 detection model. Figure 8The diagram shows the structure of the GSBottleneck module. The GS Bottleneck module is used to achieve efficient feature extraction and fusion. Its core design is a parallel-fusion path: Main path (conventional convolutional path): The input feature map is directly passed through a conventional convolutional layer to extract global and main features. Secondary path (depth-separable convolutional path): The input feature map first passes through a depth-separable convolutional layer to extract local detail features with low computational cost; then it is connected to a conventional convolutional layer for further processing. Feature fusion: The feature maps output from the above two paths are concatenated along the channel dimension. Channel shuffling: Channel shuffling is performed on the concatenated feature map to promote information interaction between features from different paths and enhance the expressive power of the features. This module significantly reduces computational complexity while maintaining feature richness by replacing some standard convolutional layers in the original network.

[0088] In yet another specific implementation, such as Figure 9 The diagram shows the structure of the VoV-GSCSP module. The VoV-GSCSP (cross-stage partial connectivity based on GSConv) module is used for efficient feature map information fusion and enhancement at different stages. Its workflow is as follows: Branch processing: The input feature map is divided into two branches for processing.

[0089] Branch 1 (Shortcut Path): The feature map passes directly through a regular convolutional layer.

[0090] Branch 2 (Feature Deepening Path): The feature map passes through a regular convolutional layer and a lightweight GSBottleneck bottleneck layer (containing residual connections) in sequence to extract deeper features.

[0091] Feature concatenation: Concatenate the output feature maps of the two branches.

[0092] Feature enhancement: The concatenated features are fused and enhanced by a 3×3 enhancement convolutional layer, and the final output is a feature map that integrates multi-level information.

[0093] The VoV-GSCSP module structure effectively integrates shallow details and deep semantic information, enhancing the model's ability to perceive the complex shapes of underwater fishing nets.

[0094] To better implement the underwater fishing net detection method based on YOLOv11 in this embodiment of the invention, based on the underwater fishing net detection method based on YOLOv11, the corresponding method is as follows: Figure 10As shown, this embodiment of the invention also provides an underwater fishing net detection device based on YOLOv11. The underwater fishing net detection device 1000 based on YOLOv11 includes: The extraction unit 1001 is used to extract features from the input underwater image to be detected through the backbone feature extraction network in the improved YOLOv11 detection model to obtain an initial feature map. The backbone feature extraction network adopts MobileNetV2 network instead of the original YOLOv11 backbone feature extraction network. The fusion unit 1002 is used to perform multi-scale feature fusion on the initial feature map through the neck network in the improved YOLOv11 detection model to obtain a fused feature map. The neck network adopts a Slim-Neck structure to replace the original YOLOv11 neck network. The detection unit 1003 is used to perform fishing net detection on the fused feature map through the detection head network in the improved YOLOv11 detection model to obtain the detection result of the fishing net target; wherein, the detection head network adopts the GSConv module, the GSBottleneck bottleneck layer and the VoV-GSCSP module to replace the original YOLOv11 detection head network.

[0095] The underwater fishing net detection device 1000 based on YOLOv11 provided in the above embodiments can realize the technical solutions described in the above embodiments of the underwater fishing net detection method based on YOLOv11. The specific implementation principles of each module or unit can be found in the corresponding content in the above embodiments of the underwater fishing net detection method based on YOLOv11, and will not be repeated here.

[0096] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0097] The above provides a detailed description of the underwater fishing net detection method and device based on YOLOv11 provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting underwater fishing nets based on YOLOv11, characterized in that, include: The improved YOLOv11 detection model uses a backbone feature extraction network to extract features from the input underwater image to be detected, resulting in an initial feature map. The backbone feature extraction network is replaced by a MobileNetV2 network instead of the original YOLOv11 backbone feature extraction network. The initial feature map is fused using the neck network in the improved YOLOv11 detection model to obtain a fused feature map. The neck network uses a Slim-Neck structure to replace the original YOLOv11 neck network. The detection head network in the improved YOLOv11 detection model is used to detect fishing nets in the fused feature map to obtain the detection result of the fishing net target; wherein, the detection head network uses the GSConv module, the GSBottleneck bottleneck layer and the VoV-GSCSP module to replace the original YOLOv11 detection head network.

2. The underwater fishing net detection method based on YOLOv11 according to claim 1, characterized in that, Before performing feature extraction on the input underwater image using the backbone feature extraction network in the improved YOLOv11 detection model, the following steps are also included: Raw underwater images are acquired using an underwater robot equipped with a monocular camera and an auxiliary light source. An improved dark channel prior enhancement algorithm is used to perform color enhancement processing on the original underwater image to generate the underwater image to be detected.

3. The underwater fishing net detection method based on YOLOv11 according to claim 2, characterized in that, The improved dark channel prior enhancement algorithm is used to perform dehazing and color enhancement processing on the original underwater image to generate the underwater image to be detected, including: The input raw underwater image is subjected to channel-adaptive dark channel estimation based on the minimum values ​​of the blue and green channels to obtain a preliminary transmittance map; Bright area extraction and background light estimation are performed on the preliminary transmittance map to obtain background light values ​​including red channel ratio correction; The preliminary transmittance map is subjected to guided filtering based on Sobel edge weights to obtain an optimized transmittance map; Based on the optimized transmittance map, relative depth calculation and adaptive gain compensation processing of the red channel are performed to obtain the compensated image, wherein the gain value of the adaptive gain compensation of the red channel is less than the adaptive upper limit related to transmittance. The compensated image is then subjected to white balance adjustment, CLAHE-based local contrast enhancement, and texture sharpening to obtain the underwater image to be detected.

4. The underwater fishing net detection method based on YOLOv11 according to claim 1, characterized in that, The MobileNetV2 network is an inverse residual structure, including a 1×1 convolutional expansion layer, a depthwise separable convolutional layer, and a 1×1 convolutional reduction layer; the feature extraction of the input underwater image is performed using the backbone feature extraction network in the improved YOLOv11 detection model to obtain an initial feature map, including: The underwater image to be detected is subjected to channel expansion processing through the 1×1 convolutional expansion layer to obtain a high-dimensional feature map. The high-dimensional feature map is processed by the depthwise separable convolutional layer to extract spatial features, thus obtaining a spatial feature map. The initial feature map is obtained by performing channel compression on the spatial feature map through the 1×1 convolutional reduction layer.

5. The underwater fishing net detection method based on YOLOv11 according to claim 1, characterized in that, The step of performing fishnet detection on the fused feature map using the detection head network in the improved YOLOv11 detection model includes: The GSConv module is used to perform feature transformation on the fused feature map to obtain a first feature map; The first feature map is input into the GSBottleneck bottleneck layer for feature compression and cross-layer information fusion to obtain the second feature map; The second feature map is input into the VoV-GSCSP module for multi-scale feature enhancement and integration to obtain an enhanced multi-scale feature map, which is then used to generate the detection result of the fishing net target.

6. The underwater fishing net detection method based on YOLOv11 according to claim 5, characterized in that, The step of using the GSConv module to perform feature transformation on the fused feature map to obtain a first feature map includes: The fused feature map is divided into first path features and second path features; Perform a standard convolution operation on the first path features and a depthwise separable convolution operation on the second path features; The first path feature after the standard convolution operation and the second path feature after the depthwise separable convolution operation are concatenated along the channel dimension to obtain the first concatenated feature. Perform a channel shuffling operation on the first spliced ​​feature to generate the first feature map.

7. The underwater fishing net detection method based on YOLOv11 according to claim 5, characterized in that, The GSBottleneck bottleneck layer includes a pre-defined GSConv module; The step of inputting the first feature map into the GSBottleneck bottleneck layer for feature compression and cross-layer information fusion to obtain the second feature map includes: The first feature map is input into the preset GSConv module in the GSBottleneck bottleneck layer for feature compression processing to obtain the basic transformation features. The basic transformation features are subjected to residual connection and channel integration processing to obtain the second feature map.

8. The underwater fishing net detection method based on YOLOv11 according to claim 1, characterized in that, The detection head network is built on the SSDLite framework, and the convolutional networks in the SSDLite framework include depthwise separable convolutional networks.

9. The underwater fishing net detection method based on YOLOv11 according to claim 5, characterized in that, The step of inputting the second feature map into the VoV-GSCSP module for multi-scale feature enhancement and integration to obtain an enhanced multi-scale feature map includes: The second feature map is divided into main path features and auxiliary path features; Perform a standard convolution operation on the main path features; The auxiliary path features are subjected to standard convolution operations and lightweight bottleneck layer processing; The main path features after standard convolution operation and the auxiliary path features after lightweight bottleneck layer processing are concatenated in the channel dimension to obtain the second concatenated feature. Perform a 3×3 convolution operation on the second spliced ​​feature to obtain an enhanced multi-scale feature map.

10. An underwater fishing net detection device based on YOLOv11, characterized in that, include: The extraction unit is used to extract features from the input underwater image to be detected through the backbone feature extraction network in the improved YOLOv11 detection model to obtain an initial feature map. The backbone feature extraction network adopts MobileNetV2 network instead of the original YOLOv11 backbone feature extraction network. The fusion unit is used to perform multi-scale feature fusion on the initial feature map through the neck network in the improved YOLOv11 detection model to obtain a fused feature map. The neck network adopts a Slim-Neck structure to replace the original YOLOv11 neck network. The detection unit is used to perform fishing net detection on the fused feature map through the detection head network in the improved YOLOv11 detection model to obtain the detection result of the fishing net target; wherein, the detection head network adopts the GSConv module, the GSBottleneck bottleneck layer and the VoV-GSCSP module to replace the original YOLOv11 detection head network.

Images