Multi-scale progressive feature enhancement method for underwater target detection

Abstract

The invention discloses a multi-scale progressive feature enhancement method for underwater target detection, and belongs to the technical field of computer vision. According to the method, based on a YOLOv8 network framework, after multi-scale features are extracted from a backbone network, a multi-scale progressive feature enhancement strategy is implemented through a neck network: a collaborative awareness and focusing module is adopted for processing a small-scale feature map so as to enhance detail and context information of a small target; performing channel and space attention weighting on the mesoscale feature map by adopting frequency domain channel space attention; a branch convolution module is used for expanding a receptive field for a large-scale feature map. And finally, after neck network fusion, a task alignment detection head outputs a detection result. According to the method, through a differential enhancement strategy, small target space details are reserved, computing resources and parameter quantities are optimized at the same time, the detection precision and robustness of the multi-scale target in the underwater complex environment are effectively improved, and a lightweight solution is provided for real-time autonomous operation of the underwater robot.

Description

Technical Field

[0001] This invention relates to computer vision and underwater real-time target detection technology, specifically to a multi-scale progressive feature enhancement method for underwater target detection. Background Technology

[0002] With the continuous advancement of marine resource development and inland river and lake construction, the demand for operational underwater robots to perform tasks in dangerous waters inaccessible to humans is becoming increasingly urgent. Underwater robots have demonstrated broad application prospects in fields such as dam inspection and maintenance, mineral sampling, underwater welding, aquaculture and fishing, and underwater security and rescue. Among them, intelligent underwater robots, capable of autonomously sensing and understanding their environment and completing complex tasks with the aid of robotic arms or specialized tools, are becoming an important development direction for "underwater workers." While traditional underwater operation methods possess certain operational capabilities, their practical applications are still limited by factors such as safety, efficiency, and environmental adaptability. In contrast, emerging intelligent detection systems, through autonomous analysis of the underwater environment, can achieve rapid detection, tracking, and decision-making regarding targets, thereby completing refined operations such as underwater sampling, demonstrating significant application value. However, current underwater detection systems still face significant technical bottlenecks in target detection.

[0003] Underwater optical imaging environments differ fundamentally from those on land, presenting numerous technical challenges. The absorption and scattering of light by water severely degrades image quality, manifesting as noticeable edge blurring and color distortion. The selective absorption of different wavelengths of light by water causes red light to attenuate completely over short distances, resulting in an overall bluish-green tint to the image and a sharp decrease in target-background contrast. Simultaneously, attenuation and forward scattering caused by suspended particles lead to image blurring and loss of detail, while backscattering creates bright spots and halos, obscuring target features. This scattering noise interference is particularly pronounced in operating environments near the seabed. Furthermore, small target detection presents a unique challenge in underwater operations. Whether it's an individual scallop during harvesting or a critical component in a black box search, these targets occupy very few pixels in the image, resulting in a severe lack of feature information. The limited information in the image makes it difficult to provide sufficient discriminative features such as shape and texture, and the downsampling operation of traditional convolutional neural networks further exacerbates the loss of small target features. Boundary blurring causes the edges of small targets to blend into the background, making precise localization difficult and leading to uneven detection performance of detectors for targets of different sizes.

[0004] In summary, there is an urgent need to develop a new underwater target detection technology that can meet the needs of autonomous underwater robot operations, and to solve core problems in underwater target detection such as edge blurring, scattering noise, and small target detection, especially for specific application scenarios such as scallop harvesting and black box search.

[0005] Against this backdrop, existing feature pyramid networks (such as FPN and PANet) employ convolutional operations to process features at different scales, failing to fully consider the inherent differences in information density and semantic level among small, medium, and large underwater targets. This homogenization leads to two limitations: first, the even distribution of limited computational resources results in insufficient enhancement of small target features while redundant computation of large target features; second, the failure to differentiate compensation for underwater-specific degradation patterns such as edge blurring and scattering noise during multi-scale feature fusion leads to the loss of spatial details for small targets. To overcome these limitations, this invention proposes a multi-scale progressive feature enhancement method. Its core lies in designing innovative branched convolutional modules and introducing a progressive optimization mechanism to effectively improve the target detection capabilities of underwater robots in complex environments, providing technical support for achieving refined underwater autonomous operations. Summary of the Invention

[0006] To address the aforementioned challenges, this invention proposes a multi-scale progressive feature enhancement method for underwater target detection: based on the YOLOv8 network framework, CPFM enhanced features are designed, and an innovative multi-scale progressive neck architecture is proposed, thereby overcoming the difficulties in existing technologies regarding the blurring of small target edges and scattering noise interference in underwater target detection.

[0007] This invention is achieved through the following technical solutions.

[0008] S1: Construct training and test sets containing labeled underwater RGB images. The training set is used for model training, and the test set is used to evaluate model performance after each training round.

[0009] S2: Input the training set into the backbone network to extract small-scale feature map C2, medium-scale feature map C3 and large-scale feature map C4.

[0010] S3: A multi-scale progressive feature enhancement strategy is used to differentiate C2, C3, and C4:

[0011] The small-scale feature map C2 is input to the Collaborative Perception and Focusing Module (CPFM) for processing to extract small-scale contextual information and generate attention weights, thereby obtaining an enhanced feature map containing rich spatial details and attention features. The CPFM includes a parallel Frequency Domain Channel Spatial Attention (FCSA) branch and a Branch Convolution Module (BCM) branch.

[0012] For the mesoscale feature map C3, input it into the frequency domain channel spatial attention FCSA to perform channel and spatial dimension attention weighting to generate the corresponding feature map.

[0013] The large-scale feature map C4 is input into the branch convolution module (BCM) for feature enhancement, thereby obtaining a feature representation with an expanded receptive field.

[0014] By employing a multi-scale progressive feature enhancement mechanism, a hierarchical and differentiated feature enhancement strategy is implemented on feature maps at different scales, thereby significantly improving the hierarchical nature of feature representation. Compared to PANet's uniformization operation, this differentiated processing strategy can more accurately enhance feature representations at each scale, significantly improving the accuracy of multi-scale object detection while maintaining computational efficiency. The multi-scale progressive feature enhancement strategy not only optimizes the utilization of computational resources but also relatively reduces the number of model parameters, thereby enhancing the overall feature representation capability.

[0015] S4: Input the multi-scale feature map processed in step S3 into the neck network for feature fusion.

[0016] S5: Input the fused features into the task alignment detection head and output the target's category and location information.

[0017] S6: Calculate the loss between the predicted value and the true value using the loss function, update the model weights using the optimizer, and iterate the training until the model converges.

[0018] S7: Use the test set to evaluate the model performance and conduct comparative experiments.

[0019] Furthermore, the CPFM module performs parallel processing of FCSA and BCM on the input feature map, and then fuses the output feature maps of the two branches element-wise. This design combines spatial awareness and local feature focusing capabilities, highlighting key features while preserving rich spatial details, thereby refining the feature representation.

[0020] For the input feature map, small target features are enhanced through two parallel branches of CPFM: Frequency Domain Channel Spatial Attention (FCSA) and Branch Convolutional Module (BCM). FCSA weights important features in both channel and spatial dimensions to highlight key information; BCM enhances feature representation through multi-branch parallel processing and residual connections. This parallel branching structure preserves rich spatial details while highlighting important spatial and channel features, improving the extraction of edge details for small underwater targets. The number of output channels in this module is equal to the number of input channels.

[0021] Furthermore, the frequency domain channel spatial attention (FCSA) includes:

[0022] Discrete Cosine Transform (DCT) is used to transform the input feature map from the spatial domain to the frequency domain.

[0023] The channel attention mechanism weights the frequency domain feature map by channel dimension and obtains the channel attention-weighted feature map by multiplying it element-wise with the original frequency domain feature map.

[0024] The spatial attention mechanism performs spatial dimension weighting on the channel-weighted feature map to obtain a channel-weighted spatial dimension-weighted feature map.

[0025] The inverse discrete cosine transform (iDCT) converts the weighted frequency domain feature map back to the spatial domain.

[0026] From the structure of FCSA, to highlight key frequency domain information from the channel dimension to the spatial dimension, important semantic information is first extracted from the input features, and then spatial information is weighted based on this. This concatenated structure, by first determining the effective feature types, can effectively reduce the computational overhead for unimportant spatial regions. Therefore, FCSA can not only simultaneously focus on the semantic content and spatial distribution of input features, but also reduce the number of redundant features, thereby optimizing overall computational efficiency.

[0027] Furthermore, the Branched Convolutional Module (BCM) includes three parallel branches:

[0028] The first branch consists of 1×1 convolution and 3×3 convolution in sequence;

[0029] The second branch includes 1×1 convolution, 1×5 convolution, 5×1 convolution and 5×5 dilated convolution in sequence;

[0030] The third branch includes 1×1 convolution, 5×1 convolution, 1×5 convolution and 5×5 dilated convolution in sequence;

[0031] The outputs of the three branches are concatenated along the channel dimension, compressed by 1×1 convolution, then fused with the original input feature map by a scaling factor, and finally activated by ReLU.

[0032] The two one-dimensional convolutional kernels in the second and third branches can process spatial information in the horizontal and vertical dimensions sequentially, thereby reducing the computational cost of each branch and achieving efficient computation with a small number of parameters. The convolutional operations with different receptive fields in these three branches help the model incorporate richer contextual information during object detection, thus more effectively capturing blurred edges.

[0033] Furthermore, the backbone network is a YOLOv8 backbone network, which outputs feature maps at three scales: C2, C3, and C4.

[0034] Furthermore, the neck network is a PAFPN structure, used for efficient fusion of multi-scale features processed by modules such as CPFM, FCSA, and BCM. This structure achieves full interaction and information integration among multi-level features through top-down semantic propagation and bottom-up detail supplementation. Specifically, the top-down path propagates high-level semantic features downwards step by step, enhancing the semantic expressive power of shallow features; while the bottom-up path fuses low-level details and localization information upwards, improving the fine-grained perception of high-level features.

[0035] Furthermore, the task alignment detection head adopts a task alignment learning mechanism, which jointly optimizes the classification and localization tasks so that the classification confidence reflects the localization quality, thereby alleviating gradient conflicts caused by inconsistent optimization objectives between the classification and localization tasks during training.

[0036] Compared with the prior art, the present invention has the following advantages and technical effects:

[0037] (1) For the multi-scale features extracted by the backbone network, a multi-scale progressive strategy is designed in the neck area. CPFM, FCSA and BCM are applied to small, medium and large features respectively to enhance them differently, preserve the spatial details of small targets and expand the receptive field of large targets, optimize computing resources and reduce the number of model parameters.

[0038] (2) Based on YOLOv8 framework training, using task alignment detection head, more balanced multi-scale detection performance is obtained in complex underwater environment, providing a lightweight solution for real-time operation of underwater robots. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall model framework of the present invention.

[0040] Figure 2 This is a structural diagram of the CPFM mentioned in this invention.

[0041] Figure 3 This is a schematic diagram of the FSCA principle mentioned in this invention.

[0042] Figure 4 This is a statistical graph of the DLMU dataset.

[0043] Figure 5 The detection effect of this invention on DLMU is shown. Detailed Implementation

[0044] To make the technical solutions, advantages and objectives of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0045] This invention proposes a multi-scale progressive feature enhancement method for underwater target detection, the overall network framework of which is as follows: Figure 1 As shown. This invention is an improvement on the YOLOv8 target detection model. YOLOv8 is the eighth generation of the "You Only Look Once" series. As an advanced single-stage anchor-free target detector, it abandons the traditional mechanism that relies on preset anchor boxes, simplifying the model structure while taking into account detection speed and accuracy, providing a good foundation for efficient target detection.

[0046] The network framework of this invention includes a backbone network, a neck network, and a head network:

[0047] 1. Backbone Network

[0048] The backbone network utilizes the existing YOLOv8 backbone structure to extract feature maps at three scales: small, medium, and large (C2, C3, C4). The resolution of these three scale feature maps decreases sequentially. C2 is a high-resolution small-scale feature map that preserves rich spatial details and is suitable for small object detection; C3 is a medium-scale feature map that balances spatial and semantic information and is used for medium-sized object detection; and C4 is a low-resolution large-scale feature map with highly abstract semantic representations and fewer spatial details, suitable for large object detection.

[0049] 2. Neck Network

[0050] The neck network is based on PAFPN and introduces a progressive enhancement strategy to optimize the utilization of computational resources and improve the overall feature representation capability. This network achieves full interaction and integration between multi-level features through top-down semantic propagation and bottom-up detail supplementation. Leveraging its bidirectional cross-connections and lateral skip structures, PAFPN effectively promotes coordination and complementarity between features at different scales, thereby enhancing semantic consistency while preserving rich spatial detail information.

[0051] 3. Head Network

[0052] The head network employs a task-aligned head to mitigate gradient conflicts caused by inconsistent optimization objectives between classification and localization tasks during training. This head includes a class branch and a location regression branch: the former outputs the confidence scores for each class, while the latter predicts the distances (l, r, t, b) from the current pixel to the four sides of the ground truth bounding box. Unlike traditional head-based heads that independently optimize classification and regression, the task-aligned head constructs a joint optimization objective through TAL (TaskAlignment Learning), enabling classification confidence scores to dynamically reflect localization quality, thereby achieving effective alignment between classification and localization tasks.

[0053] The specific implementation process of the method of the present invention is as follows:

[0054] S1: The underwater RGB image dataset publicly available from DLMU was used. This dataset includes labeled training and test sets. The training set was used for model training, and the test set was used to evaluate model performance after each training round. The model was trained on an NVIDIA TITAN RTX (24576 MiB), using Python version 3.10.0 and torch version 2.9.0+cu126. After each training round, the model parameters were updated and evaluated on the test set to obtain performance metrics such as accuracy, regression accuracy, and AP at various scales, thereby enabling effective monitoring of the training process.

[0055] The DLMU dataset features low image continuity and high data diversity, with data sources from more than 10 different sea areas, providing reliable support for the rapid development of underwater detection research.

[0056] Currently, the number of publicly available underwater detection datasets is limited, and their overall quality is not ideal. The main problems are high image continuity and low data diversity. These issues prevent trained models from reliably applying to complex and variable underwater environments. Furthermore, these datasets typically collect images from a single sea area, resulting in highly consistent backgrounds. This limits the trained detectors to handling only a single scenario within the dataset. To address these problems, this embodiment uses the DLMU2024 dataset (DLMU for short). DLMU contains 2500 underwater samples and four underwater categories: sea cucumbers, sea urchins, scallops, and starfish. These categories primarily target aquatic organisms common in marine ranches. The training and test sets are divided in an 8:2 ratio, with 2000 images in the training set and 500 images in the test set.

[0057] In the DLMU dataset, small, medium, and large targets account for 40.42%, 49.95%, and 9.63%, respectively. Figure 4 As shown, among small targets, sea urchins (Echinus) accounted for the highest proportion at 58.3%. The other three categories were distributed in a pyramid shape across the three scales, meaning that small and large targets accounted for roughly the same proportion, and the proportion of small targets was smaller compared to medium targets, which is helpful for evaluating the model's detection performance for small targets.

[0058] S2: Input the training set into the backbone network and extract small-scale feature map C2, medium-scale feature map C3 and large-scale feature map C4;

[0059] S3: A multi-scale progressive feature enhancement strategy is used to differentiate C2, C3, and C4:

[0060] (1) The small-scale feature map C2 is input to the Collaborative Perception and Focusing Module (CPFM) for processing. The CPFM includes a parallel Frequency Domain Channel Spatial Attention (FCSA) branch and a Branch Convolution Module (BCM) branch. The structure diagram is shown below. Figure 2 As shown, this design adaptively enhances the feature response of key areas while fully preserving spatial details, thereby achieving a more refined feature representation.

[0061] CPFM simultaneously feeds the input feature map X into two parallel branches: Frequency Domain Channel Spatial Attention (FCSA) and Branch Convolutional Module (BCM). FCSA converts the input into a frequency domain feature map, then weights the spatial and channel information of the frequency domain feature map to output the attention feature map X. attn The Branched Convolutional Module (BCM) extracts features from different receptive fields through multi-branch parallel processing and residual connections, outputting an enhanced feature map X. en Finally, the attention feature map X attn and enhanced feature map X en Feature fusion is performed by adding elements one by one. This parallel branching structure not only preserves rich spatial details but also highlights important spatial and channel features in the foreground information, improving the feature extraction of edge details of small underwater targets.

[0062] Therefore, CPFM not only overcomes the limitation of FCSA, which can only weight feature information and cannot substantially enhance feature representation capabilities, but also further improves the feature representation ability. In addition, CPFM compensates for the noise and redundant information problems caused by the lack of a global attention mechanism in BCM by integrating collaborative perception and focusing mechanisms.

[0063] (2) For the mesoscale feature map C3, input it into the frequency domain channel spatial attention FCSA to perform channel and spatial dimension attention weighting. The module structure is as follows: Figure 3 As shown, FCSA not only allows the network model to focus on the more important frequency domain information in the feature map, but also suppresses unimportant frequency domain information, thereby extracting features more efficiently.

[0064] Discrete Cosine Transform (DCT) transforms the input spatial domain feature map X t Convert to frequency domain feature map X f This transforms the feature map from a spatial domain signal to a frequency domain signal:

[0065]

[0066] Channel attention processes the input feature map through the following steps:

[0067] First, global average pooling and global max pooling operations are performed on the input feature map to obtain important semantic information, resulting in two feature vectors of size B×C×1×1. This process is represented as:

[0068]

[0069]

[0070] Subsequently, the two feature vectors are flattened to a B×C size and then input into a multilayer perceptron (MLP) for weight sharing. This MLP contains two linear layers: the first compresses the number of channels from C to C / ratio, and the second linear layer restores the number of channels from C / ratio to C. The feature vectors output by these two MLPs are then element-wise summed and activated using a sigmoid function to obtain the channel attention weights.

[0071] Spatial attention processes the channel-weighted feature map through the following steps: Max pooling and average pooling are performed on the feature map along the channel dimension to extract important spatial information, resulting in two feature maps of size B×1×H×W. This process is represented as follows:

[0072]

[0073]

[0074] After pooling, the two feature maps are concatenated along the channel dimension. Then, a 7×7 convolutional layer compresses the output channel number of the concatenated feature map to 1. Next, the compressed feature map is activated using the sigmoid function to generate spatial attention weights. These weights are then multiplied element-wise with the channel-attention-weighted feature map to generate the output frequency domain feature map Y. f Inverse discrete cosine transform is used to obtain the output feature map Y. t :

[0075]

[0076] like Figure 2 The overall module structure of FCSA, as shown, involves Channel Attention and Spatial Attention weighting the input feature maps and then connecting them sequentially. This connection method allows the network to accurately focus on the channel and location information of the feature maps. Furthermore, FCSA adds only a few computational parameters and computational cost, yet delivers significant performance improvements. Its lightweight structure guides the network to concentrate limited computational resources on the feature channels and spatial locations of the information.

[0077] (3) Input the large-scale feature map C4 into the branch convolution module BCM for feature enhancement.

[0078] like Figure 2As shown, BCM processes the input feature map X through three parallel branches: the first branch performs 1×1 and 3×3 convolutions sequentially; the second branch performs 1×1, 1×5, 5×1, and 5×5 dilated convolutions sequentially; and the third branch performs 1×1, 5×1, 1×5, and 5×5 dilated convolutions sequentially. These convolution operations with different receptive fields in these three branches help the model incorporate richer contextual information during object detection, thus more effectively capturing blurred edges. The two one-dimensional convolutional kernels in the second and third branches can process spatial information in the horizontal and vertical dimensions sequentially, thereby reducing the computational cost of each branch and achieving efficient computation with a small number of parameters.

[0079] Then, the feature maps output from the three branches are concatenated along the channel dimension, and the concatenated feature map X1 is compressed by a 1×1 convolutional layer to obtain X2 = CBS(X1).

[0080] To preserve the original feature information while incorporating the rich feature information after multi-branch processing, the original feature map shortcut(X) after the compressed feature map X2 is shortened is weighted and fused using a scaling factor: X3 = X2 × scale + shortcut(X), where scale is a preset scaling factor. Finally, the fused feature map X3 is processed by the ReLU activation function, and the output feature map Y = ReLU(X3) is obtained.

[0081] S4: Input the multi-scale feature map processed in step S3 into the neck network for feature fusion.

[0082] S5: Input the fused features into the task alignment detection head and output the target's category and location information.

[0083] S6: Calculate the loss between the predicted and true values ​​using a loss function, update the model weights using an optimizer, and iterate the training until the model converges. During this process, for each true value, use the TOP-K algorithm to select the K highest-matching predicted boxes from its candidate boxes, labeling them as positive samples, and using the others as negative samples to fit the background. During training, K is set to 10, and the training run is performed for 100 epochs.

[0084] S7: Use the test set for testing. The information obtained from the test is first filtered out to remove all prediction boxes with low confidence. Then, redundant boxes are removed by non-maximum suppression to reduce the amount of computation. The accuracy P, regression rate R, AP, AP50, APS and other indicators of the model for targets at various scales are obtained, and then the final model after training is obtained.

[0085] The following comparative experiments verify the detection capability and lightweight effectiveness of this invention in complex underwater environments. To comprehensively evaluate the multi-scale progressive feature enhancement method proposed in this invention, we selected RetinaNet, YOLOv5, YOLOv8, and AMSP-UOD detection algorithms as benchmarks on two lightweight scales. Given the lack of a dedicated evaluation system in the field of underwater target detection, this study adopts the common COCO evaluation protocol, quantifying classification accuracy and localization precision using AP (IoU=0.50~0.95), AP50 (IoU=0.50), and AP75 (IoU=0.75) indicators. Simultaneously, APS, APM, and APL indicators are used to measure the model's detection capability for targets at different scales.

[0086] The visualization results of this invention on the DLMU test set are as follows: Figure 5 As shown in the figure, the experimental images generally present challenges such as dense small targets, multiple targets mixed together, and color cast and dispersion noise, which significantly increases the detection difficulty. Specifically, the starfish and sea cucumbers in Figure (a) that are highly similar to the background, and the densely distributed sea cucumbers in Figure (b), both pose a severe test to the model's discrimination ability. The detection results show that the multi-scale progressive enhancement method can accurately identify target categories and locate their boundaries, verifying its target detection capability in complex underwater scenes such as small targets, scattering noise interference, and blurred edges.

[0087] The table below shows the metrics for different models: based on the DLMU dataset, trained in Scratch for 100 rounds.

[0088]

[0089] The bolded data in the table represents the best performance among models of the same size, while the underlined data represents the second best performance among models of the same size. As shown in the table, for nano(n) scale, our invention (Ours) achieves the best performance in Recall, AP, AP50, AP75, APS, APM, and APL, with improvements of 0.3%, 0.3%, 0.2%, 1.2%, 0.1%, 0.1%, and 0.3% respectively compared to the second best models. Compared to the traditional YOLO algorithm and the AMSP-UOD underwater target detection network based on AMSP-VConv, our invention achieves better performance across all metrics.

[0090] For the small(s) scale model, the regression rates Recall, AP, AP50, AP75, APM, and APL achieved the best performance, at 72.3%, 45.3%, 77.1%, 50.2%, 38.2%, and 50%, respectively, representing improvements of 0.6%, 1.9%, 0.9%, 0.6%, 0.4%, and 2.2% compared to the suboptimal model. The proposed method achieved suboptimal APS at 4.5%, a 0.9% improvement over the YOLOv8 model.

[0091] At the nano scale, the precision index is slightly lower than some comparative models. This phenomenon can be explained from the perspective of the trade-offs in model design: First, this invention allocates more model capacity to recall optimization (Recall reaches the optimal 69%), and captures more potential small targets by relaxing the classification confidence threshold. This will objectively introduce some false detections, which is a common phenomenon in the precision-recall trade-off. Second, the CPFM module adopts a parallel structure of BCM and FCSA, which may enhance the response intensity of background interference areas while enhancing the details of small targets.

[0092] YOLOv8's nano and small scales differ significantly in maximum channel count and number of repeating layers. The nano-scale backbone network has a maximum of only 256 channels, while the small scale increases this to 512. Furthermore, the small scale's C2 and C3 feature map channels are 128 and 256 respectively, significantly higher than the nano scale's 64 and 128. This structural difference implies that, at the small scale, the backbone network extracts higher-dimensional basic features, resulting in stronger semantic carrying capacity for the neck network. This provides more information redundancy for the parallel dual-branch structure of the CPFM module, enabling the attention mechanism to more accurately distinguish between the target and the background, thus achieving comprehensive improvements in metrics such as APS and APM.

[0093] In tests of both scales, the proposed method showed a significant improvement in the detection capability of small targets compared to the original YOLOv8 model.

[0094] The table below shows various indicators for different models.

[0095]

[0096] From a category perspective, the baseline model (YOLOv8) and the proposed method (Ours) exhibit consistent performance characteristics: the detection accuracy for sea urchins (Echinus) and starfish (Starfish) is significantly higher than that for sea cucumbers (Holothurian) and scallops (Scallop). Among the two models shown in the table, scallops have the lowest detection accuracy: AP is less than 30% for both, and AP50 is less than 50%. This is mainly due to the differences in the characteristics of the targets themselves: sea urchins and starfish typically have regular shapes and relatively clear textures or colors that contrast with the background, making them relatively easy to detect; while sea cucumbers and scallops have irregular shapes, surface textures highly similar to the seabed environment, and are often covered in mud and sand, making them difficult samples that require a higher level of feature discrimination capability from the model.

[0097] The proposed method significantly improves the detection accuracy for sea urchins and starfish, and for scallops, it improves the accuracy by 0.3%. This indicates that the proposed method, through targeted improvements, effectively enhances the model's feature extraction and discrimination capabilities in scenarios with blurred edges of small targets and high noise interference. As a result, it exhibits higher detection accuracy and generalization performance when facing difficult targets that are similar to the background and subject to severe interference.

[0098] The table below shows the neck ablation experiments, mainly comparing parameter quantities and GFLOPs.

[0099]

[0100] In the table, "-" indicates that the module is not used, and "√" indicates that the module is used. If only one "√" is marked, it means that the information transmitted by C2, C3, and C4 is processed using only this module. The table shows that the number of parameters and GFLOPs using only CPFM, FCSA, and BCM are 12.68, 11.82, 12.64 and 32.8, 31, 32.7 respectively, both increasing compared to the baseline model's 11.14 and 28.6. The model using only FCSA shows the smallest increase. The multi-scale progressive feature enhancement method proposed in this invention balances the computational cost of the three modules, reducing the number of parameters and GFLOPs by 0.63 and 1.5 respectively compared to the model using only CPFM. This demonstrates the effectiveness of the multi-scale progressive enhancement method proposed in this invention in reducing the number of parameters.

Claims

1. A multi-scale progressive feature enhancement method for underwater target detection, characterized in that, Includes the following steps: S1: Construct training and testing sets containing labeled underwater RGB images; S2: Input the training set into the backbone network and extract small-scale feature map C2, medium-scale feature map C3 and large-scale feature map C4; S3: A multi-scale progressive feature enhancement strategy is used to differentiate C2, C3, and C4: - The small-scale feature map C2 is input to the Collaborative Perception and Focusing Module (CPFM) for processing. The CPFM includes a parallel Frequency Domain Channel Spatial Attention (FCSA) branch and a Branch Convolution Module (BCM) branch. - Input the mesoscale feature map C3 into the frequency domain channel spatial attention FCSA to perform channel and spatial dimension attention weighting; - Input the large-scale feature map C4 into the branch convolutional module BCM for feature enhancement; S4: Input the multi-scale feature map processed in step S3 into the neck network for feature fusion; S5: Input the fused features into the task alignment detection head and output the target's category and location information; S6: Calculate the loss between the predicted value and the true value using the loss function, update the model weights using the optimizer, and iterate the training until the model converges. S7: Use the test set to evaluate model performance.

2. The method of claim 1, wherein, The CPFM module performs FCSA and BCM parallel processing on the input feature map simultaneously, and then adds and fuses the output feature maps of the two branches element by element.

3. The method according to claim 1 or 2, characterized in that, The frequency domain channel spatial attention (FCSA) includes: Discrete Cosine Transform (DCT) is used to transform the input feature map from the spatial domain to the frequency domain. Channel attention mechanism, which weights the frequency domain feature map by channel dimension; Spatial attention mechanism, which performs spatial dimension weighting on the channel-weighted feature map; The inverse discrete cosine transform (iDCT) converts the weighted frequency domain feature map back to the spatial domain.

4. The method of claim 1, wherein, The Branched Convolutional Module (BCM) includes three parallel branches: The first branch consists of 1×1 convolution and 3×3 convolution in sequence; The second branch includes 1×1 convolution, 1×5 convolution, 5×1 convolution and 5×5 dilated convolution in sequence; The third branch includes 1×1 convolution, 5×1 convolution, 1×5 convolution and 5×5 dilated convolution in sequence; The outputs of the three branches are concatenated along the channel dimension, compressed by 1×1 convolution, then fused with the original input feature map by a scaling factor, and finally activated by ReLU.

5. The method according to claim 1, characterized in that, The backbone network is the YOLOv8 backbone network, which outputs feature maps at three scales: C2, C3, and C4.

6. The method according to claim 1, characterized in that, The neck network is a PAFPN structure, used for top-down and bottom-up feature fusion of multi-scale features processed by CPFM, FCSA, and BCM.

7. The method according to claim 1, characterized in that, The task alignment detection head adopts a task alignment learning mechanism, which jointly optimizes classification and localization tasks so that classification confidence reflects localization quality.

Images