A method for intelligent identification and deployment of multiple targets under sea surface based on lightweight YOLOv8
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEBEI JIESHUANG AIRLINES TECHNOLOGY CO LTD
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-30
Smart Images
Figure CN122313007A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent recognition technology for multiple targets underwater, and in particular to a method for intelligent recognition and deployment of multiple targets underwater based on lightweight YOLOv8. Background Technology
[0002] With the increasing demand for marine resource development and seabed operations, the automatic identification of underwater targets using underwater vision systems has become an important technical means in aquaculture monitoring, underwater robotic operations, and marine resource surveys. By deploying visual recognition systems on underwater robots or autonomous underwater vehicles, the automatic identification and positioning of underwater targets such as sea cucumbers, sea urchins, and scallops can be achieved, thereby improving the efficiency of marine operations and reducing labor costs.
[0003] In existing technologies, underwater target recognition typically employs general target detection networks, such as the YOLO series networks or other convolutional neural network structures, which are trained on underwater images to achieve target detection. However, the underwater imaging environment differs significantly from the terrestrial imaging environment. The absorption and scattering of light by water leads to phenomena such as red light attenuation, blue and green light enhancement, and scattering by suspended particles, resulting in severe color cast, reduced contrast, and blurred details in underwater images.
[0004] Existing technologies typically combine underwater image enhancement algorithms with target detection algorithms in a simple chain, such as first enhancing the underwater image and then inputting the enhanced image into a target detection network for recognition. However, this separate processing approach cannot guarantee feature consistency between the image enhancement process and the target detection feature extraction process. The enhancement algorithm may alter the color distribution or texture structure of the original image, thereby interfering with the feature representation of the target detection network and leading to a decrease in the stability of the detection model in complex underwater environments.
[0005] In the subsurface environment, water turbidity varies significantly with changes in suspended particle concentration, lighting conditions, and water depth. When environmental turbidity is high, shallow visual features often contain a large amount of high-frequency noise generated by the scattering of suspended particles. Existing target detection networks typically employ fixed-structure feature pyramids or path aggregation structures when performing multi-scale feature fusion, lacking adaptive feature filtering mechanisms for underwater turbid environments. This easily introduces noise information during the feature fusion stage, thereby affecting the accuracy of target detection. Summary of the Invention
[0006] One objective of this invention is to propose a method for intelligent identification and deployment of multiple targets underwater based on lightweight YOLOv8. This invention effectively reduces the impact of suspended particle scattering, background flickering texture, and local optical noise on target detection, and improves the robustness of underwater multi-target detection in turbid water environments.
[0007] A method for intelligent identification and deployment of multiple targets underwater based on lightweight YOLOv8 according to an embodiment of the present invention includes:
[0008] Collect underwater optical images of underwater operational scenarios to construct an underwater multi-target recognition image dataset;
[0009] The color-aware lightweight YOLOv8 detection network was trained end-to-end using an underwater multi-target recognition image dataset, and the network parameters were updated to obtain the trained color-aware lightweight YOLOv8 detection model.
[0010] During the inference phase, the underwater optical images acquired in real time are input into the color adaptive enhancement network of the trained color-aware lightweight YOLOv8 detection model to generate enhanced underwater optical images and global color bias vectors.
[0011] A color-aware backbone network is used to perform hierarchical feature extraction on enhanced underwater optical images, outputting shallow high-resolution feature maps that preserve the geometric details of small targets, and deep low-resolution feature maps that contain semantic information of targets.
[0012] Channel attention weights are generated using global color bias vectors, and channel weighted modulation is performed on shallow high-resolution feature maps and deep low-resolution feature maps respectively to generate shallow color bias embedding feature maps and deep color bias embedding feature maps.
[0013] A turbidity-gated feature pyramid is constructed, and a channel-level suppression mask is used to adaptively weight and filter the shallow color-biased embedding feature map, retaining only the fusion path of the deep semantic features, to generate a set of noise-resistant fusion feature maps.
[0014] Dynamic upsampling is performed on the fused feature map set to output a reconstructed feature map. The reconstructed feature map is then input into the dynamic perception detection head to generate target class probability and target bounding box prediction results.
[0015] The trained color-perception lightweight YOLOv8 detection model is subjected to structured pruning, parameter quantization and graph model transformation to generate an edge inference model. The edge inference model is then deployed on the embedded computing platform of an underwater robot or autonomous underwater vehicle to achieve real-time identification and localization of multiple targets underwater.
[0016] Optionally, the underwater multi-target recognition image dataset includes sea cucumber target image samples, sea urchin target image samples, scallop target image samples, and corresponding annotation information.
[0017] Optionally, the step of using an underwater multi-target recognition image dataset to perform end-to-end training and network parameter updates on a color-aware lightweight YOLOv8 detection network includes:
[0018] Construct a training sample set for an underwater multi-target recognition image dataset;
[0019] Each underwater optical image sample in the training sample set is input into the color adaptive enhancement network. During the training phase, a training enhanced underwater optical image and a training global color bias vector are generated. Automatic color equalization processing is performed on the same underwater optical image sample to generate a reference enhanced underwater optical image.
[0020] In the color-perception lightweight YOLOv8 detection network, a weight-sharing relationship is established between the encoder and the color-perception backbone network. The encoder of the color adaptive enhancement network with the same network topology as the first four feature extraction stages of the color-perception backbone network is constructed.
[0021] The enhanced underwater optical image is input into the color-aware backbone network that satisfies the weight sharing relationship, and the output is a set of multi-scale backbone feature maps. At the same time, the original underwater optical image sample is input into the encoder of the color adaptive enhancement network to obtain the encoded feature set.
[0022] For shared hierarchical indexes, a color representation consistency constraint loss is constructed to ensure that the channel statistical distribution of the encoded feature set and the multi-scale backbone feature map set remains consistent at the same level.
[0023] The enhanced underwater optical image is input into the color-aware lightweight YOLOv8 detection network to obtain a set of predicted bounding boxes.
[0024] Based on the predicted bounding box set and the true bounding box set, calculate the minimum point distance intersection-union ratio regression loss;
[0025] Based on the enhanced underwater optical image and the reference enhanced underwater optical image, the spectral consistency error loss corresponding to the spectral consistency error loss function is calculated.
[0026] The joint loss function is obtained by weighting and summing the minimum point distance intersection-union regression loss, the spectral consistency error loss, and the color representation consistency constraint loss. The parameter set of the color-aware lightweight YOLOv8 detection network is iteratively updated using the gradient descent algorithm until the joint loss function converges, thus obtaining the trained color-aware lightweight YOLOv8 detection model.
[0027] Optionally, the step of using a color-aware backbone network to perform hierarchical feature extraction on the enhanced underwater optical image, and outputting shallow high-resolution feature maps that preserve the geometric details of small targets, includes:
[0028] The enhanced underwater optical image corresponding to the nth underwater optical image sample is input into the color-perceived backbone network to establish a hierarchical convolutional feature extraction process for the color-perceived backbone network, and the enhanced underwater optical image is defined as the 0th level input feature map.
[0029] The first-level front-end convolutional feature extraction structure and the second-level front-end convolutional feature extraction structure are sequentially executed on the 0th-level input feature map to obtain a shallow high-resolution feature map.
[0030] The third and fourth levels of deep convolutional feature extraction structures are sequentially applied to the shallow high-resolution feature map to obtain the deep low-resolution feature map.
[0031] Optionally, the step of generating channel attention weights using a global color bias vector and performing channel-weighted modulation on the shallow high-resolution feature map and the deep low-resolution feature map respectively includes:
[0032] The shallow channel attention weight vector corresponding to the shallow high-resolution feature map is generated using the global color bias vector.
[0033] The deep channel attention weight vector corresponding to the deep low-resolution feature map is generated using the global color bias vector.
[0034] By using the shallow channel attention weight vector and the deep channel attention weight vector respectively, channel-wise weighted modulation is performed on the shallow high-resolution feature map and the deep low-resolution feature map to generate the shallow color-biased embedding feature map and the deep color-biased embedding feature map.
[0035] Optionally, the construction of the turbidity-gated feature pyramid, using channel-level suppression masks to adaptively weight and filter the shallow color-biased embedding feature maps, retaining only the fusion path of deep semantic features, includes:
[0036] The channel-level suppression mask corresponding to the nth underwater optical image sample is generated using the environmental turbidity coefficient;
[0037] An adaptive weighted filtering method is used to perform channel-level suppression masking on the shallow color-biased embedding feature map to obtain a shallow noise-resistant feature map.
[0038] Based on the comparison between the environmental turbidity coefficient and the preset threshold, a turbidity gating feature pyramid is constructed.
[0039] The turbidity-gated feature pyramid is output as the set of noise-resistant fusion feature maps, and the set of noise-resistant fusion feature maps is defined as consisting of the first noise-resistant fusion feature map and the second noise-resistant fusion feature map.
[0040] Optionally, the turbidity-gated feature pyramid construction rules include:
[0041] When the environmental turbidity coefficient is not higher than the preset threshold, the turbidity gating feature pyramid contains both a shallow noise reduction feature map and a deep color bias embedding feature map.
[0042] When the environmental turbidity coefficient is higher than the preset threshold, the turbidity gating feature pyramid only contains the deep color bias embedding feature map, and the shallow fusion path is set to zero response input to indicate the shallow fusion path is closed.
[0043] Optionally, the step of performing dynamic upsampling processing on the fused feature map set, outputting a reconstructed feature map, and inputting the reconstructed feature map into the dynamic perception detection head includes:
[0044] Input the set of noise-resistant fusion feature maps corresponding to the nth underwater optical image sample into the dynamic upsampling operator;
[0045] Using a dynamic upsampling operator, dynamic upsampling processing is performed on the second noise-resistant fusion feature map to generate an upsampled feature map with the same spatial resolution as the first noise-resistant fusion feature map.
[0046] The upsampled feature map is channel-aligned and element-wise fused with the first noise-resistant fusion feature map to obtain the reconstructed feature map.
[0047] The reconstructed feature map is input into the dynamic perception detection head, and scale-aware attention is used to perform scale-dimension weighting on the reconstructed feature map to obtain the scale-enhanced feature map.
[0048] Spatial location weighting is performed on the scale-enhanced feature map to obtain the spatially enhanced feature map;
[0049] Task-dimensional weighted processing is performed on the spatial augmentation feature map using task-aware attention to obtain the task-augmented feature map;
[0050] Based on the task-enhanced feature map, the target class probability and target bounding box prediction results are generated.
[0051] Optionally, the calculation rules for the reconstructed feature map include:
[0052] When the first noise-resistant fusion feature map is in a non-zero response state, the upsampled feature map and the first noise-resistant fusion feature map are added element-wise at the same spatial position and the same channel position to obtain the reconstructed feature map;
[0053] When the first noise-resistant fusion feature map is in a zero-response input state, the upsampled feature map is directly output as the reconstructed feature map.
[0054] The beneficial effects of this invention are:
[0055] (1) This invention constructs a weight-sharing training mechanism between the color adaptive enhancement network and the color perception backbone network, so that the underwater image enhancement process and the target detection feature extraction process can be jointly optimized in the same feature space, thereby achieving deep collaboration between image enhancement and target detection. By establishing a weight-sharing relationship between the encoder convolutional layer of the color adaptive enhancement network and the first four convolutional layers of the color perception backbone network in the end-to-end training stage, and by constraining the channel statistical distribution of the encoded feature set and the backbone feature set through color representation consistency constraint loss, the underwater imaging degradation feature learning process and the target semantic feature learning process are aligned in a unified feature space. This enables the enhanced underwater optical image to not only compensate for the color shift caused by underwater red light attenuation and blue-green scattering, but also to maintain the key geometric details of the sea cucumber target outline boundary, the sea urchin target spiny structure, and the scallop target shell texture, thus significantly improving the expressive ability of small underwater targets in the feature space.
[0056] (2) This invention proposes a turbidity-sensing gated feature pyramid structure driven by the environmental turbidity coefficient to realize an adaptive feature fusion mechanism for underwater turbid environments. By using the global color bias vector to calculate the environmental turbidity coefficient and constructing a channel-level suppression mask, adaptive weighted filtering is performed on each channel in the shallow color bias embedded feature map. When the environmental turbidity is high, the high-frequency noise channel in the shallow features is automatically suppressed. At the same time, the shallow fusion path is closed through the gating mechanism, and only the fusion path of the deep semantic features is retained. When the environmental turbidity is low, both the shallow geometric details and the deep semantic features are retained. Through the turbidity-driven dynamic feature fusion strategy, the feature fusion structure can automatically adjust the contribution ratio of shallow and deep features according to the underwater environmental conditions, effectively reducing the impact of suspended particle scattering, background flickering texture and local optical noise on target detection, and improving the robustness of multi-target detection under the sea surface in turbid water environments. Attached Figure Description
[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0058] Figure 1 This is a flowchart of a method for intelligent identification and deployment of multiple targets under the sea surface based on lightweight YOLOv8 proposed in this invention;
[0059] Figure 2 This invention presents the overall framework of a color-perception lightweight YOLOv8 detection network in a method for intelligent identification and deployment of multiple targets underwater based on lightweight YOLOv8. Detailed Implementation
[0060] Example 1: Reference Figures 1-2A method for intelligent identification and deployment of multiple targets underwater based on lightweight YOLOv8, comprising:
[0061] Collect underwater optical images of underwater operational scenarios to construct an underwater multi-target recognition image dataset;
[0062] In this embodiment, the underwater multi-target recognition image dataset includes sea cucumber target image samples, sea urchin target image samples, scallop target image samples, and corresponding annotation information.
[0063] The color-aware lightweight YOLOv8 detection network was trained end-to-end using an underwater multi-target recognition image dataset, and the network parameters were updated to obtain the trained color-aware lightweight YOLOv8 detection model.
[0064] In this embodiment, an underwater multi-target recognition image dataset is used to perform end-to-end training and network parameter updates on a color-aware lightweight YOLOv8 detection network, including:
[0065] Construct a training sample set for an underwater multi-target recognition image dataset;
[0066] In Example 1, the training sample set consists of several training sample pairs. Each training sample pair consists of an underwater optical image sample and a corresponding set of ground truth bounding boxes. The underwater optical image sample represents the original underwater optical image collected in the underwater working environment. The set of ground truth bounding boxes represents the spatial annotation information of all underwater targets in the underwater optical image sample. All underwater optical image samples in the training sample set are subjected to pixel intensity normalization processing to keep the pixel intensity within the range of zero to one.
[0067] Each underwater optical image sample in the training sample set is input into the color adaptive enhancement network. During the training phase, a training enhanced underwater optical image and a training global color bias vector are generated. Automatic color equalization processing is performed on the same underwater optical image sample to generate a reference enhanced underwater optical image.
[0068] In Example 1, multi-layer convolutional feature extraction is performed on underwater optical image samples in the encoder of the color adaptive enhancement network to obtain underwater imaging degradation characterization feature maps. In the decoder of the color adaptive enhancement network, a color compensation image is reconstructed based on the underwater imaging degradation characterization feature maps. An enhanced underwater optical image is generated by pixel-by-pixel superposition of the color compensation image and the underwater optical image samples.
[0069] Global average pooling is performed on the encoder output features of the color adaptive enhancement network to obtain a three-dimensional color statistical vector, and a linear mapping is performed to obtain a global color bias vector. The global color bias vector consists of red channel offset, green channel offset, and blue channel offset, which is used to describe the comprehensive color bias caused by red light attenuation, green light scattering, and blue light dominance in the underwater optical imaging process.
[0070] The color compensation intensity is obtained by comparing and accumulating the brightness difference between each pixel in the underwater optical image sample and its neighboring pixels. Based on the color compensation intensity, pixel-by-pixel color adjustment is performed on the underwater optical image sample to generate a reference enhanced underwater optical image.
[0071] In the color-perception lightweight YOLOv8 detection network, a weight-sharing relationship is established between the encoder and the color-perception backbone network. The encoder of the color adaptive enhancement network with the same network topology as the first four feature extraction stages of the color-perception backbone network is constructed.
[0072] In Example 1, the weights of the encoder convolutional layer of the color adaptive enhancement network are exactly the same as the weights of the first four convolutional layers of the color perception backbone network. The weights of the encoder convolutional layer are used to extract imaging degradation features in underwater optical images, and the weights of the first four convolutional layers of the color perception backbone network are used to extract target semantic features under the sea surface. Through the weight sharing relationship, the degradation feature learning process and the target semantic feature learning process are aligned in the same feature space.
[0073] The enhanced underwater optical image is input into the color-aware backbone network that satisfies the weight sharing relationship, and the output is a set of multi-scale backbone feature maps. At the same time, the original underwater optical image sample is input into the encoder of the color adaptive enhancement network to obtain the encoded feature set.
[0074] In Example 1, feature extraction of enhanced underwater optical images is performed sequentially through a four-level convolutional feature extraction structure in the color perception backbone network. Each level of the convolutional feature extraction structure includes convolution operation, nonlinear activation operation, and spatial downsampling operation. At the same time, channel attention weights are generated using the global color bias vector, and the channel attention weights are embedded into the output features of each level of the convolutional feature extraction structure to perform channel weighted modulation, thereby realizing the dynamic calibration of the target semantic features by physical color bias information. By progressively reducing the spatial resolution and progressively increasing the number of feature channels, the texture structure information and semantic structure information of the target under the sea surface in the enhanced underwater optical image are expressed in feature spaces at different scales, resulting in a multi-scale backbone feature map set, which is used to represent the semantic feature map of the target under the sea surface calibrated by color bias information at different spatial resolution scales.
[0075] In the encoder of the color adaptive enhancement network, convolutional feature extraction is performed on the original underwater optical image samples using the same convolutional weights as the first four convolutional layers of the color perception backbone network. After each convolutional layer, nonlinear activation and spatial downsampling operations are performed to map the underwater optical imaging degradation information in the original underwater optical image samples layer by layer into a high-dimensional feature space, resulting in a set of encoded features. This set represents the underwater imaging degradation feature map extracted from the original underwater optical image samples under the same convolutional weights.
[0076] For shared hierarchical indexes, a color representation consistency constraint loss is constructed to ensure that the channel statistical distribution of the encoded feature set and the multi-scale backbone feature map set remains consistent at the same level.
[0077] In Example 1, each multi-scale backbone feature map is subjected to global average pooling to obtain channel statistical vectors. The absolute difference between the encoded feature channel statistical vector and the backbone feature channel statistical vector is calculated. All channel absolute differences are summed. The statistical differences between all training samples and all shared layers are averaged to obtain the color representation consistency constraint loss.
[0078] The enhanced underwater optical image is input into the color-aware lightweight YOLOv8 detection network to obtain a set of predicted bounding boxes.
[0079] In Example 1, the predicted bounding box set represents the positions of all underwater target bounding boxes predicted by the color-aware lightweight YOLOv8 detection network in the current underwater optical image sample, and is matched with the real bounding boxes in the set of real bounding boxes.
[0080] Based on the predicted bounding box set and the true bounding box set, calculate the minimum point distance intersection-union ratio regression loss;
[0081] ;
[0082] ;
[0083] in, This represents the internal intersection-union ratio calculated based on the overlapping regions within the bounding box. This represents the Euclidean distance between the top-left corner of the predicted bounding box and the ground truth bounding box. This represents the Euclidean distance between the bottom right corner of the predicted bounding box and the ground truth bounding box. Indicates the actual bounding box width. Indicates the actual bounding box height. This indicates to prevent the denominator from being calculated. This represents matching bounding box pairs, where N represents the number of training samples. This represents the number of underwater targets contained in the nth underwater optical image sample. This represents the minimum point distance intersection ratio regression loss.
[0084] Based on the enhanced underwater optical image and the reference enhanced underwater optical image, the spectral consistency error loss corresponding to the spectral consistency error loss function is calculated.
[0085] In Example 1, the spectral consistency error loss is used to constrain the consistency of the enhanced underwater optical image and the reference enhanced underwater optical image in the pixel-level spectral distribution. It is obtained by calculating the absolute difference between the enhanced underwater optical image and the reference enhanced underwater optical image at each pixel position, summing the absolute differences of all pixel positions and all color channels, and averaging over all training samples.
[0086] The joint loss function is obtained by weighting and summing the minimum point distance intersection-union regression loss, the spectral consistency error loss, and the color representation consistency constraint loss. The parameter set of the color-aware lightweight YOLOv8 detection network is iteratively updated using the gradient descent algorithm until the joint loss function converges, thus obtaining the trained color-aware lightweight YOLOv8 detection model.
[0087] In Example 1, both the spectral consistency error loss and the color characterization consistency constraint loss are adaptively weighted according to the environmental turbidity coefficient. When the environmental turbidity coefficient is large, the weight of the spectral consistency error loss is increased, and when the environmental turbidity coefficient is small, the weight of the color characterization consistency constraint loss is increased. The environmental turbidity coefficient is obtained by calculating the square of the difference between the blue channel offset and the red channel offset, and the square of the difference between the green channel offset and the red channel offset, and then performing a square root operation after summing the two squared values. The larger the environmental turbidity coefficient, the stronger the blue and green scattering and the more obvious the red light attenuation in the underwater environment.
[0088] This implementation introduces a joint loss function consisting of minimum point distance intersection-over-union regression loss, spectral consistency error loss, and color representation consistency constraint loss during the end-to-end training of the lightweight YOLOv8 color perception detection model. The minimum point distance intersection-over-union regression loss improves the accuracy of small-scale target bounding box localization underwater. The spectral consistency error loss constrains the consistency of the enhanced underwater optical image's pixel-level spectral distribution with the automatic color equalization enhancement result, reducing the interference of visual distortion caused by water scattering and color shift on the detection model training. The color representation consistency constraint loss adjusts the statistical distribution of the encoded feature set and the multi-scale backbone feature map set within the same level of channels. The system enables collaborative learning of underwater optical imaging degradation features and underwater target semantic features within the same feature space. Simultaneously, based on the turbidity characteristics of the underwater environment, an environmental turbidity coefficient is introduced to adaptively adjust the weights of spectral consistency error loss and color representation consistency constraint loss. This allows the lightweight YOLOv8 color perception detection model to focus more on color compensation capabilities in highly turbid environments and on semantic feature consistency in low-turbid environments. While maintaining the low computational complexity of the lightweight YOLOv8 detection network, the system improves the stability and detection accuracy of intelligent multi-target recognition underwater, achieving reliable identification of small underwater targets such as sea cucumbers, sea urchins, and scallops in complex marine optical environments.
[0089] During the inference phase, the underwater optical images acquired in real time are input into the color adaptive enhancement network of the trained color-aware lightweight YOLOv8 detection model to generate enhanced underwater optical images and global color bias vectors.
[0090] A color-aware backbone network is used to perform hierarchical feature extraction on enhanced underwater optical images, outputting shallow high-resolution feature maps that preserve the geometric details of small targets, and deep low-resolution feature maps that contain semantic information of targets.
[0091] In this embodiment, a color-aware backbone network is used to perform layered feature extraction on the enhanced underwater optical image, outputting shallow high-resolution feature maps that preserve the geometric details of small targets, including:
[0092] The enhanced underwater optical image corresponding to the nth underwater optical image sample is input into the color-perceived backbone network to establish a hierarchical convolutional feature extraction process for the color-perceived backbone network, and the enhanced underwater optical image is defined as the 0th level input feature map.
[0093] The first-level front-end convolutional feature extraction structure and the second-level front-end convolutional feature extraction structure are sequentially executed on the 0th-level input feature map to obtain a shallow high-resolution feature map.
[0094] The first-level front-end convolutional feature extraction structure obtains the first-level convolutional response feature map by performing a convolution operation between the first-level input feature map and the weight parameters of the first-level convolutional layer. The first-level activation feature map is obtained by performing a non-linear activation operation on the first-level convolutional response feature map. Finally, the first-level backbone feature map is obtained by performing a spatial downsampling operation on the first-level activation feature map.
[0095] The second-level front-end convolutional feature extraction structure obtains the second-level convolutional response feature map by performing a convolution operation between the first-level backbone feature map and the weight parameters of the second-level convolutional layer. The second-level activation feature map is obtained by performing a nonlinear activation operation on the second-level convolutional response feature map, and the second-level backbone feature map is obtained by performing a spatial downsampling operation on the second-level activation feature map. The second-level backbone feature map is defined as a shallow high-resolution feature map, which is used to express the geometric details of small targets under the sea surface, such as the outline boundary of sea cucumber targets, the spiny edge of sea urchin targets, and the shell texture of scallop targets in enhanced underwater optical images.
[0096] The third and fourth levels of deep convolutional feature extraction structures are sequentially applied to the shallow high-resolution feature map to obtain the deep low-resolution feature map.
[0097] In Example 1, the solution for deep low-resolution feature maps is the same as that for shallow high-resolution feature maps. Deep low-resolution feature maps are used to express the semantic structure information of the categories of targets under the sea surface, including the elongated semantic structure of sea cucumber targets, the central clustering semantic structure of sea urchin targets, and the fan-shaped shell semantic structure of scallop targets.
[0098] Channel attention weights are generated using global color bias vectors, and channel weighted modulation is performed on shallow high-resolution feature maps and deep low-resolution feature maps respectively to generate shallow color bias embedding feature maps and deep color bias embedding feature maps.
[0099] In this embodiment, channel attention weights are generated using a global color bias vector, and channel-weighted modulation is performed on shallow high-resolution feature maps and deep low-resolution feature maps respectively, including:
[0100] The shallow channel attention weight vector corresponding to the shallow high-resolution feature map is generated using the global color bias vector.
[0101] In Example 1, the global color bias vector is input into the shallow weight mapping matrix for linear mapping operation to obtain the shallow linear mapping vector. The shallow bias vector is superimposed on the shallow linear mapping vector to obtain the shallow bias mapping vector. The sigmoid activation function is performed on the shallow bias mapping vector to obtain the shallow channel attention weight vector. Each channel weight in the shallow channel attention weight vector corresponds to the response intensity of the corresponding channel in the shallow high-resolution feature map to the geometric details of small targets under the sea surface.
[0102] The shallow weight mapping matrix represents the shallow weight mapping matrix that maps the global color bias vector to the shallow high-resolution feature map channel space, and the shallow bias vector represents the shallow bias vector in the shallow channel mapping process.
[0103] The deep channel attention weight vector corresponding to the deep low-resolution feature map is generated using the global color bias vector.
[0104] The calculation of the deep channel attention weight vector is the same as that of the shallow channel attention weight vector. Each channel weight in the deep channel attention weight vector corresponds to the response intensity of the corresponding channel in the deep low-resolution feature map to the semantic structure information of the target category below the sea surface.
[0105] By using the shallow channel attention weight vector and the deep channel attention weight vector respectively, channel-wise weighted modulation is performed on the shallow high-resolution feature map and the deep low-resolution feature map to generate the shallow color-biased embedding feature map and the deep color-biased embedding feature map.
[0106] A turbidity-gated feature pyramid is constructed, and a channel-level suppression mask is used to adaptively weight and filter the shallow color-biased embedding feature map, retaining only the fusion path of the deep semantic features, to generate a set of noise-resistant fusion feature maps.
[0107] In this embodiment, a turbidity-gated feature pyramid is constructed, and channel-level suppression masks are used to adaptively weight and filter the shallow color-biased embedding feature maps, retaining only the fusion path of deep semantic features, including:
[0108] The channel-level suppression mask corresponding to the nth underwater optical image sample is generated using the environmental turbidity coefficient;
[0109] In Example 1, the channel-level suppression mask is used to describe the retention weight of each channel in the shallow color bias embedding feature map under the current turbidity conditions of the sea surface environment, so that the high-frequency noise channels that are significantly affected by the scattering of suspended particles are adaptively suppressed in the feature fusion stage.
[0110] ;
[0111] in, Indicates the first The channel-level suppression mask corresponding to the underwater optical image sample is in the th... Mask weights on each channel This represents the Sigmoid activation function. Indicates the first The turbidity response scaling factor for each channel is used to adjust the environmental turbidity coefficient. The magnitude of the effect on the channel inhibition intensity This represents a preset threshold used to distinguish between low and high turbidity conditions in the subsurface environment. Indicates the first The mask offset corresponding to the first channel is used to adjust the first channel's mask offset. The basic retention level of each channel This represents the channel index of the shallow color bias embedding feature map.
[0112] An adaptive weighted filtering method is used to perform channel-level suppression masking on the shallow color-biased embedding feature map to obtain a shallow noise-resistant feature map.
[0113] In Example 1, each channel in the shallow color bias embedding feature map is processed separately. The channel-level suppression mask weight of the corresponding channel is multiplied with the shallow color bias embedding feature response of the corresponding channel to obtain the filtering result of the channel. The filtering results of all channels are combined to obtain the shallow noise reduction feature map.
[0114] When the shallow fusion path is preserved, as the environmental turbidity coefficient increases, the mask weight of the corresponding high-frequency noise channel in the channel-level suppression mask adaptively decreases, so that the channel response caused by the reflection of suspended particles under the sea surface, background flicker texture and local scattering noise in the shallow color bias embedding feature map is continuously suppressed.
[0115] Based on the comparison between the environmental turbidity coefficient and the preset threshold, a turbidity gating feature pyramid is constructed.
[0116] In this embodiment, the rules for constructing the turbidity-gated feature pyramid include:
[0117] When the environmental turbidity coefficient is not higher than the preset threshold, the turbidity gating feature pyramid contains both a shallow noise reduction feature map and a deep color bias embedding feature map.
[0118] When the environmental turbidity coefficient is higher than the preset threshold, the turbidity gating feature pyramid only contains the deep color bias embedding feature map, and the shallow fusion path is set to zero response input to indicate the shallow fusion path is closed.
[0119] The turbidity-gated feature pyramid is output as the set of noise-resistant fusion feature maps, and the set of noise-resistant fusion feature maps is defined as consisting of the first noise-resistant fusion feature map and the second noise-resistant fusion feature map.
[0120] In Example 1, when the environmental turbidity coefficient is not higher than the preset threshold, the first noise reduction fusion feature map is a shallow noise reduction feature map, and the second noise reduction fusion feature map is a deep color shift embedding feature map.
[0121] When the environmental turbidity coefficient is higher than the preset threshold, the first noise-resistant fusion feature map is a zero-response input, and the second noise-resistant fusion feature map is a deep color-biased embedding feature map.
[0122] The noise-resistant fusion feature map set directly inherits the gating results of the turbidity-gated feature pyramid. When the turbidity of the underwater environment is high, the shallow fusion path is closed, and only the fusion path corresponding to the deep color-bias embedding feature map is retained. When the turbidity of the underwater environment is low, the fusion paths corresponding to the shallow noise-resistant feature map and the deep color-bias embedding feature map are retained at the same time.
[0123] This implementation introduces an environmental turbidity coefficient calculation mechanism based on a global color bias vector, and combines channel-level suppression masks with a turbidity-gated feature pyramid structure to achieve adaptive feature fusion control for complex optical imaging environments beneath the sea surface. It utilizes the environmental turbidity coefficient to dynamically suppress high-frequency noise channels in shallow high-resolution feature maps, and automatically closes the shallow fusion path when environmental turbidity is high, retaining only deep semantic features for feature fusion. This effectively reduces the interference of underwater suspended particle scattering, background flickering textures, and local optical noise on the target recognition process. Simultaneously, when environmental turbidity is low, it retains shallow noise-resistant feature maps, allowing the outline boundaries of sea cucumber targets, the spiny edges of sea urchin targets, and the geometric details of small targets such as scallop shell textures to fully participate in the feature fusion process. This enables the color-perceptive lightweight YOLOv8 detection model to adaptively adjust the feature fusion path according to the degree of underwater environmental turbidity, improving the stability and noise resistance of multi-target recognition beneath the sea surface while ensuring the model's lightweight deployment capability.
[0124] Dynamic upsampling is performed on the fused feature map set to output a reconstructed feature map. The reconstructed feature map is then input into the dynamic perception detection head to generate target class probability and target bounding box prediction results.
[0125] In this embodiment, dynamic upsampling processing is performed on the fused feature map set to output a reconstructed feature map. The reconstructed feature map is then input into the dynamic perception detection head, including:
[0126] Input the set of noise-resistant fusion feature maps corresponding to the nth underwater optical image sample into the dynamic upsampling operator;
[0127] Using a dynamic upsampling operator, dynamic upsampling processing is performed on the second noise-resistant fusion feature map to generate an upsampled feature map with the same spatial resolution as the first noise-resistant fusion feature map.
[0128] In Example 1, channel compression and pointwise convolution operations are performed on the second noise-resistant fusion feature map to reduce the number of channels to a preset recombination dimension, and the output is rearranged in the spatial dimension. Normalization is performed using the Softmax function to generate dynamic local recombination convolution kernels corresponding to each spatial location.
[0129] ;
[0130] in, Indicates the first Dynamic local reconstruction convolution kernels corresponding to each underwater optical image sample The sample index representing the underwater optical image sample, PixelShuffle This represents the pixel rearrangement operator. Indicates the kernel size as Convolution operation, Indicates the first The second noise-resistant fusion feature map corresponds to each underwater optical image sample. For each spatial location in the second noise-resistant fusion feature map, a dynamic local recombination convolution kernel is used to weight and recombine the neighboring feature responses to obtain an upsampled feature map:
[0131] ;
[0132] in, Indicates the first The upsampled feature map corresponding to each underwater optical image sample is located in the spatial position First Characteristic response values on each channel, This represents the spatial coordinate index of the upsampled feature map in the horizontal direction. This represents the spatial coordinate index of the upsampled feature map in the vertical direction. The channel index of the feature map. This represents the neighborhood offset index of the dynamically locally reorganized convolutional kernel in the horizontal direction. This represents the neighborhood offset index of the dynamically locally recombined convolutional kernel in the vertical direction. This represents the neighborhood radius of the dynamically locally recombined convolutional kernel. This represents the dynamic upsampling ratio, used to indicate the scale mapping relationship between the spatial coordinates of the upsampled feature map and the spatial coordinates of the second noise-resistant fusion feature map. This indicates the floor function.
[0133] The upsampled feature map is channel-aligned and element-wise fused with the first noise-resistant fusion feature map to obtain the reconstructed feature map.
[0134] In this embodiment, the calculation rules for reconstructing the feature map include:
[0135] When the first noise-resistant fusion feature map is in a non-zero response state, the upsampled feature map and the first noise-resistant fusion feature map are added element-wise at the same spatial position and the same channel position to obtain the reconstructed feature map;
[0136] When the first noise-resistant fusion feature map is in a zero-response input state, the upsampled feature map is directly output as the reconstructed feature map.
[0137] In Example 1, when the turbidity of the underwater environment is low, the first noise-resistant fusion feature map retains the geometric details of small underwater targets corresponding to the outline boundary of the sea cucumber target, the spiny edge of the sea urchin target, and the shell texture of the scallop target, and together with the target semantic structure information in the upsampled feature map, it forms a reconstructed feature map; when the turbidity of the underwater environment is high, the first noise-resistant fusion feature map is a zero-response input, and the reconstructed feature map is composed only of the upsampled feature map.
[0138] The reconstructed feature map is input into the dynamic perception detection head, and scale-aware attention is used to perform scale-dimension weighting on the reconstructed feature map to obtain the scale-enhanced feature map.
[0139] Spatial location weighting is performed on the scale-enhanced feature map to obtain the spatially enhanced feature map;
[0140] Task-dimensional weighted processing is performed on the spatial augmentation feature map using task-aware attention to obtain the task-augmented feature map;
[0141] In Example 1, the reconstructed feature map is input into the dynamic perception detection head. The dynamic perception detection head sequentially performs scale-aware attention processing, spatial awareness attention processing, and task awareness attention processing. The scale-aware attention is used to weight the feature responses of the reconstructed feature map at different scale levels to obtain a scale-enhanced feature map.
[0142] By applying spatial awareness attention weights to the corresponding spatial locations of the scale-enhanced feature map, the spatial response of the areas where sea cucumber, sea urchin, and scallop targets are located is enhanced, while the spatial response of the background water area, the suspended particle scattering area, and the targetless area is suppressed, thus obtaining the spatially enhanced feature map.
[0143] Task-enhanced feature maps are obtained by weighting the contribution of each channel in the spatial augmentation feature map to the target category discrimination task and the target bounding box regression task using task-aware attention.
[0144] Based on the task-enhanced feature map, the target class probability and target bounding box prediction results are generated.
[0145] In Example 1, the task enhancement feature map is input into the classification branch mapping network. Linear mapping calculation is performed on the feature vector corresponding to each spatial location in the task enhancement feature map to obtain the response value of each candidate target in different categories. Normalization probability mapping processing is performed on the response value of each category to obtain the target category probability, which is used to represent the probability distribution of each candidate target in the current underwater optical image sample belonging to sea cucumber, sea urchin or scallop target.
[0146] The task-enhanced feature map is input into the regression branch mapping network. Boundary box parameter regression calculation is performed on the feature vector corresponding to each spatial location in the task-enhanced feature map. By jointly predicting the target center position offset, target width parameter, and target height parameter, the target bounding box prediction result is obtained, which is used to represent the spatial position prediction result of each candidate target in the current underwater optical image sample.
[0147] The trained color-perception lightweight YOLOv8 detection model is subjected to structured pruning, parameter quantization and graph model transformation to generate an edge inference model. The edge inference model is then deployed on the embedded computing platform of an underwater robot or autonomous underwater vehicle to achieve real-time identification and localization of multiple targets underwater.
[0148] Example 2: In a certain underwater aquaculture monitoring mission, this invention was deployed to automatically identify underwater biological targets such as sea cucumbers, sea urchins, and scallops. The system consists of underwater optical cameras and an embedded computing platform. During the inspection process, it continuously acquires underwater images and performs target identification tasks in real time. In one continuous inspection cycle, the system acquired a total of 21,436 underwater optical images, with a uniform image resolution of 1280×720 pixels. A training sample set was constructed through manual annotation, and each image recorded the target category and corresponding bounding box position, including 15,260 sea cucumber targets, 10,935 sea urchin targets, and 13,742 scallop targets.
[0149] During the training phase, the system inputs training samples into the automatic color equalization and co-enhancement module. In an underwater optical image sample numbered Sample-08734, the system detected that the average value of the red channel was 0.24, the average value of the green channel was 0.47, and the average value of the blue channel was 0.63. Based on the color statistics, the system calculates a global color shift vector, obtaining a red channel offset of -0.36, a green channel offset of 0.12, and a blue channel offset of 0.41. The system then performs color compensation processing on the input image based on this color shift vector. After compensation, the average brightness of the image increases from 0.41 to 0.56, the contrast increases from 0.28 to 0.49, and the color shift index decreases from 0.53 to 0.17. The system records the enhanced image as an enhanced underwater optical image sample and continues to input the image into the feature extraction network.
[0150] The enhanced image is fed into a color-aware backbone network for multi-layer feature extraction. In the first convolutional output stage, the system generates a feature map with dimensions of 640×360×64. In the second convolutional output stage, a shallow, high-resolution feature map with dimensions of 320×180×128 is generated. In this feature map, the system detects a sea cucumber target contour with a response value of 0.74, while the average response value for the background region is 0.21. The system records shallow feature maps that preserve significant geometric details, including the sea cucumber's body surface outline, sea urchin spines, and scallop shell edge texture.
[0151] The network continues with deep feature extraction, generating a feature map of size 160×90×256 at the output stage of the third convolutional layer, and a deep semantic feature map of size 80×45×512 at the output stage of the fourth convolutional layer. The system records that in this deep feature map, the average semantic activation value for sea cucumbers is 0.82, for sea urchins it is 0.79, and for scallops it is 0.76, while the average response value for the background region is only 0.18.
[0152] During the feature modulation stage, the system generates channel attention weights based on the global color bias vector. In the shallow feature map, the system detected a weight of 0.93 for channel 12, 0.88 for channel 25, and only 0.21 for channel 41. The system determined that channel 41 is a high-frequency scattering noise channel, and therefore suppressed it during the feature fusion stage. The system recorded that the high-frequency noise energy in the shallow feature map decreased from 0.62 to 0.27.
[0153] During the turbidity gating stage, the system calculates the environmental turbidity coefficient based on the color shift vector. In the current sample, the calculated turbidity coefficient is 0.68, while the system's preset threshold is 0.55. Since the turbidity coefficient exceeds the threshold, the system automatically triggers a shallow path suppression strategy. The system records the weights of channel 1 (0.81), channel 2 (0.76), channel 3 (0.12), and channel 4 (0.09) in the channel-level suppression mask. The system determines channels 3 and 4 to be high-frequency noise channels and therefore performs strong suppression on them. In the turbidity gating feature pyramid generated by the system in the sample, only the deep color shift embedding feature map is retained, and the shallow fusion path is set to zero response input.
[0154] During the dynamic upsampling phase, the system inputs deep features into the dynamic upsampling module. The system generates a local dynamic convolutional kernel at spatial location (154, 201), recording the center weight of the kernel as 0.26 and the weights at adjacent locations as 0.12 and 0.19, respectively. The system uses the convolutional kernel to locally reconstruct the deep features, restoring the feature map size from 80×45×512 to 320×180×256 after upsampling. The system records a significant enhancement in the target semantic structure in the reconstructed feature map, with the average response value of the sea cucumber target region increasing from 0.72 to 0.85.
[0155] The reconstructed feature map is input into the dynamic perception detection head module. In the scale-aware attention processing stage, the scale weight of the detected small target region is increased by 1.27 times; in the spatial perception attention processing stage, the spatial response of the regions where sea cucumbers, sea urchins, and scallops are located is enhanced to 0.91, while the response of the background region is reduced to 0.14; in the task-aware attention processing stage, the classification channel weight is increased to 0.83, and the bounding box regression channel weight is increased to 0.79.
[0156] The system ultimately generates target category probabilities and bounding box prediction results. In Sample-08734, the system identified 6 targets: 3 sea cucumbers, 1 sea urchin, and 2 scallops. The system recorded the confidence scores for the sea cucumber targets as 0.94, 0.91, and 0.89, the sea urchin target as 0.86, and the scallop targets as 0.93 and 0.90, respectively. The system calculated the average intersection-union ratio (IUU) between the predicted and ground truth bounding boxes to be 0.91.
[0157] During the testing phase, the system performed recognition experiments on 1880 test images and was compared with the traditional YOLOv8 method. The traditional method uses conventional image enhancement and a standard YOLOv8 detection network. Under the same test data conditions, the traditional method has an average detection precision of 0.823, an average recall of 0.79, a false positive rate of 0.13, and an average inference time of 33 milliseconds. However, when using the method of this invention, the system's average detection precision is improved to 0.907, the average recall rate is improved to 0.91, the false positive rate is reduced to 0.06, and the average inference time is reduced to 24 milliseconds.
[0158] During a continuous inspection, the system identified a total of 1823 sea cucumbers, 1042 sea urchins, and 1365 scallops. Manual verification showed that the overall identification accuracy of the method described in this invention reached 91.8%, while the traditional method only achieved 82.4%. In scenarios with high water turbidity, the traditional method's detection accuracy dropped to 0.742, while the method described in this invention maintained a detection accuracy of 0.861.
[0159] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for intelligent identification and deployment of multiple targets underwater based on lightweight YOLOv8, characterized in that, include: Collect underwater optical images of underwater operational scenarios to construct an underwater multi-target recognition image dataset; The color-aware lightweight YOLOv8 detection network was trained end-to-end using an underwater multi-target recognition image dataset, and the network parameters were updated to obtain the trained color-aware lightweight YOLOv8 detection model. During the inference phase, the underwater optical images acquired in real time are input into the color adaptive enhancement network of the trained color-aware lightweight YOLOv8 detection model to generate enhanced underwater optical images and global color bias vectors. A color-aware backbone network is used to perform hierarchical feature extraction on enhanced underwater optical images, outputting shallow high-resolution feature maps that preserve the geometric details of small targets, and deep low-resolution feature maps that contain semantic information of targets. Channel attention weights are generated using global color bias vectors, and channel weighted modulation is performed on shallow high-resolution feature maps and deep low-resolution feature maps respectively to generate shallow color bias embedding feature maps and deep color bias embedding feature maps. A turbidity-gated feature pyramid is constructed, and a channel-level suppression mask is used to adaptively weight and filter the shallow color-biased embedding feature map, retaining only the fusion path of the deep semantic features, to generate a set of noise-resistant fusion feature maps. Dynamic upsampling is performed on the fused feature map set to output a reconstructed feature map. The reconstructed feature map is then input into the dynamic perception detection head to generate target class probability and target bounding box prediction results. The trained color-perception lightweight YOLOv8 detection model is subjected to structured pruning, parameter quantization and graph model transformation to generate an edge inference model. The edge inference model is then deployed on the embedded computing platform of an underwater robot or autonomous underwater vehicle to achieve real-time identification and localization of multiple targets underwater.
2. The method for intelligent identification and deployment of multiple targets underwater based on lightweight YOLOv8 according to claim 1, characterized in that, The underwater multi-target recognition image dataset includes sea cucumber target image samples, sea urchin target image samples, scallop target image samples, and corresponding annotation information.
3. The method for intelligent identification and deployment of multiple underwater targets based on lightweight YOLOv8 according to claim 1, characterized in that, The process of training a lightweight YOLOv8 color-aware detection network end-to-end using an underwater multi-target recognition image dataset and updating its network parameters includes: Construct a training sample set for an underwater multi-target recognition image dataset; Each underwater optical image sample in the training sample set is input into the color adaptive enhancement network. During the training phase, a training enhanced underwater optical image and a training global color bias vector are generated. Automatic color equalization processing is performed on the same underwater optical image sample to generate a reference enhanced underwater optical image. In the color-perception lightweight YOLOv8 detection network, a weight-sharing relationship is established between the encoder and the color-perception backbone network. An encoder with the same network topology as the first four feature extraction stages of the color-perception backbone network is constructed. The enhanced underwater optical image is input into the color-aware backbone network that satisfies the weight sharing relationship, and the output is a set of multi-scale backbone feature maps. At the same time, the original underwater optical image sample is input into the encoder of the color adaptive enhancement network to obtain the encoded feature set. For shared hierarchical indexes, a color representation consistency constraint loss is constructed to ensure that the channel statistical distribution of the encoded feature set and the multi-scale backbone feature map set remains consistent at the same level. The enhanced underwater optical image is input into the color-aware lightweight YOLOv8 detection network to obtain a set of predicted bounding boxes. Based on the predicted bounding box set and the true bounding box set, calculate the minimum point distance intersection-union ratio regression loss; Based on the enhanced underwater optical image and the reference enhanced underwater optical image, the spectral consistency error loss corresponding to the spectral consistency error loss function is calculated. The joint loss function is obtained by weighted summation of the minimum point distance intersection-union regression loss, the spectral consistency error loss, and the color representation consistency constraint loss. The parameter set of the color-aware lightweight YOLOv8 detection network is iteratively updated using the gradient descent algorithm until the joint loss function converges, thus obtaining the trained color-aware lightweight YOLOv8 detection model.
4. The method for intelligent identification and deployment of multiple underwater targets based on lightweight YOLOv8 according to claim 3, characterized in that, The method utilizes a color-aware backbone network to perform hierarchical feature extraction on enhanced underwater optical images, outputting shallow, high-resolution feature maps that preserve the geometric details of small targets, including: The enhanced underwater optical image corresponding to the nth underwater optical image sample is input into the color-perceived backbone network to establish a hierarchical convolutional feature extraction process for the color-perceived backbone network, and the enhanced underwater optical image is defined as the 0th level input feature map. The first-level front-end convolutional feature extraction structure and the second-level front-end convolutional feature extraction structure are sequentially executed on the 0th-level input feature map to obtain a shallow high-resolution feature map. The shallow high-resolution feature map is processed sequentially using the third-level deep convolutional feature extraction structure and the fourth-level deep convolutional feature extraction structure to obtain the deep low-resolution feature map.
5. The method for intelligent identification and deployment of multiple underwater targets based on lightweight YOLOv8 according to claim 4, characterized in that, The step of generating channel attention weights using a global color bias vector and performing channel-weighted modulation on shallow high-resolution feature maps and deep low-resolution feature maps respectively includes: The shallow channel attention weight vector corresponding to the shallow high-resolution feature map is generated using the global color bias vector. The deep channel attention weight vector corresponding to the deep low-resolution feature map is generated using the global color bias vector. By using the shallow channel attention weight vector and the deep channel attention weight vector respectively, channel-wise weighted modulation is performed on the shallow high-resolution feature map and the deep low-resolution feature map to generate the shallow color-biased embedding feature map and the deep color-biased embedding feature map.
6. The method for intelligent identification and deployment of multiple underwater targets based on lightweight YOLOv8 according to claim 5, characterized in that, The construction of the turbidity-gated feature pyramid, which uses channel-level suppression masks to adaptively weight and filter shallow color-biased embedding feature maps, retaining only the fusion path of deep semantic features, includes: The channel-level suppression mask corresponding to the nth underwater optical image sample is generated using the environmental turbidity coefficient; An adaptive weighted filtering method is used to perform channel-level suppression masking on the shallow color-biased embedding feature map to obtain a shallow noise-resistant feature map. Based on the comparison between the environmental turbidity coefficient and the preset threshold, a turbidity gating feature pyramid is constructed. The turbidity-gated feature pyramid is output as the set of noise-resistant fusion feature maps, and the set of noise-resistant fusion feature maps is defined as consisting of the first noise-resistant fusion feature map and the second noise-resistant fusion feature map.
7. The method for intelligent identification and deployment of multiple underwater targets based on lightweight YOLOv8 according to claim 6, characterized in that, The turbidity-gated feature pyramid construction rules include: When the environmental turbidity coefficient is not higher than the preset threshold, the turbidity gating feature pyramid contains both a shallow noise-resistant feature map and a deep color bias embedding feature map. When the environmental turbidity coefficient is higher than the preset threshold, the turbidity gating feature pyramid only contains the deep color bias embedding feature map, and the shallow fusion path is set to zero response input to indicate the shallow fusion path is closed.
8. The method for intelligent identification and deployment of multiple underwater targets based on lightweight YOLOv8 according to claim 6, characterized in that, The process of performing dynamic upsampling on the fused feature map set, outputting a reconstructed feature map, and inputting the reconstructed feature map into the dynamic perception detection head includes: Input the set of noise-resistant fusion feature maps corresponding to the nth underwater optical image sample into the dynamic upsampling operator; Using a dynamic upsampling operator, dynamic upsampling processing is performed on the second noise-resistant fusion feature map to generate an upsampled feature map with the same spatial resolution as the first noise-resistant fusion feature map. The upsampled feature map is channel-aligned and element-wise fused with the first noise-resistant fusion feature map to obtain the reconstructed feature map. The reconstructed feature map is input into the dynamic perception detection head, and scale-aware attention is used to perform scale-dimension weighting on the reconstructed feature map to obtain the scale-enhanced feature map. Spatial location weighting is performed on the scale-enhanced feature map to obtain the spatially enhanced feature map; Task-dimensional weighted processing is performed on the spatial augmentation feature map using task-aware attention to obtain the task-augmented feature map; Based on the task-enhanced feature map, the target class probability and target bounding box prediction results are generated.
9. A method for intelligent identification and deployment of multiple underwater targets based on lightweight YOLOv8 according to claim 8, characterized in that, The calculation rules for the reconstructed feature map include: When the first noise-resistant fusion feature map is in a non-zero response state, the upsampled feature map and the first noise-resistant fusion feature map are added element-wise at the same spatial position and the same channel position to obtain the reconstructed feature map; When the first noise-resistant fusion feature map is in a zero-response input state, the upsampled feature map is directly output as the reconstructed feature map.