A method and system for identifying small target ships at sea in a scattering environment based on YOLO-SOD

Abstract

The application discloses a small target ship identification method and system based on YOLO-SOD in a scattering environment, and relates to the field of visual target detection. At present, in the scattering condition, the small target ship on the sea has the problems of low pixel ratio, weak texture, semantic smoothing and easy mis-detection and missed detection in the sea-sky line neighborhood. The application comprises the following steps: acquiring a sea image and calculating a scattering intensity index; constructing a YOLO-SOD network comprising a scattering perception dynamic convolution main stem, a learnable weight bidirectional feature fusion neck, a P2-P5 multi-scale detection head group and a sea-sky line uncertain band double-head differentiated detection sub-network; adopting a label allocation strategy combining target scale and scattering intensity and a small target weighted loss function containing a center offset constraint during training; dynamically adjusting the convolution kernel weight, adaptively fusing the features, constructing the uncertain band for double-branch joint detection during inference, and outputting the boundary box coordinates and confidence of the small target ship. The technical scheme effectively improves the detection recall rate and positioning accuracy of the small target ship in the scattering environment.

Description

Technical Field

[0001] This invention relates to the field of visual target detection, and in particular to a method and system for identifying small marine targets in a scattering environment based on YOLO-SOD. Background Technology

[0002] Currently, the detection of small targets at long distances at sea is characterized by significant weak features and strong interference. Weak features manifest in the low pixel ratio of the target, lack of texture detail, and blurred boundary information. Strong interference arises from scattering conditions such as sea fog, salt spray, wave spray, and backlight diffusion, which reduce image contrast, smooth target textures, and enhance background noise. Simultaneously, numerous false targets resembling small vessels, such as wave crests, highlights, and cloud edges, exist in the vicinity of the sea horizon. The combined effect of these two factors significantly reduces the separability of the target from the background, leading to widespread problems in existing detection systems, including high false positive rates, high missed detection rates, and low positioning accuracy.

[0003] In existing technologies, ship target detection under complex sea conditions mainly adopts a "front-end image enhancement + general target detector" approach. This involves first improving the visual quality of the input image through algorithms such as dehazing, enhancement, and noise reduction, and then inputting the processed image into a standard general detection network such as YOLO. This approach is effective in near-to-medium range, weak scattering scenarios, but it has significant limitations in strong scattering, long-range, small target detection scenarios. Front-end image enhancement can only improve the appearance quality of the image and cannot fundamentally solve the problem of insufficient network ability to extract weak features of small targets. In addition, some enhancement algorithms will introduce new noise or artifacts, which will increase the difficulty of detection. The backbone of the general YOLO network uses a fixed convolution kernel and a fixed downsampling rhythm, which cannot adaptively adjust the feature extraction strategy according to different scattering intensities. Early downsampling will cause the remaining texture information of small targets to be lost prematurely. The neck feature fusion structure adopts a fixed-weight bidirectional fusion method, which cannot automatically adjust the fusion ratio of shallow detail features and deep semantic features according to the data distribution, resulting in shallow details related to small targets being excessively masked by deep smooth semantics. The detection end uses a uniform detection head and a fixed confidence threshold, which makes it impossible to perform differentiated modeling of the sea-line neighborhood, a region with a high density of false targets. It is difficult to simultaneously take into account both the overall recall rate and the false detection suppression of the sea-line region. The label assignment strategy during the training phase only considers the target scale and does not take into account the scene scattering intensity. This results in the supervision signal of small target samples in strong scattering scenes being diluted by medium and large target samples. The loss function is also not specifically optimized for the localization jitter problem of small targets at a distance.

[0004] Therefore, there is an urgent need to develop an efficient detection method that can adapt to scattering environments and is specifically designed for detecting small, long-range targets at sea. Summary of the Invention

[0005] The technical problem to be solved and the technical task proposed by this invention is to improve and refine existing technical solutions, and to provide a method and system for identifying small maritime targets under scattering environments based on YOLO-SOD, so as to achieve high-precision and robust detection of small maritime targets under scattering environments. To this end, this invention adopts the following technical solution.

[0006] Firstly, a method for identifying small marine targets in a scattering environment based on YOLO-SOD includes the following steps: 1) Acquire images or video frames at sea and calculate their scattering intensity index; 2) Construct the YOLO-SOD network, which includes a scattering-sensing dynamic convolutional backbone network, a bidirectional feature fusion neck network with learnable weights, a P2~P5 multi-scale detection head group, and a dual-head differential detection sub-network for uncertain sea-line areas; 3) Train the YOLO-SOD network by inputting the training samples and their scattering intensity indices into the YOLO-SOD network, using a scale-aware label allocation strategy that combines target scale and scattering intensity indices to perform positive sample matching, and using a small target weighted joint loss function with center offset constraints to optimize the network parameters. 4) Input the image to be identified and its scattering intensity index into the trained network: The backbone network dynamically adjusts the weights of its first two convolutional kernels based on the scattering intensity index and outputs multi-scale feature maps; the neck network performs adaptive weighted fusion of the feature maps and outputs P2 high-resolution feature maps and P3~P5 feature maps for detecting small targets at a distance; the detection head group detects each feature map to obtain preliminary features; the sea-line uncertain band dual-head differential detection sub-network dynamically constructs uncertain bands based on the sea-line estimated position, temporal drift variance and basic bandwidth, and uses a dedicated branch within the uncertain band and a main sea surface branch outside the band to jointly infer the preliminary features, outputting the bounding box coordinates and confidence of the small target ship.

[0007] This technical solution calculates the scattering intensity index and inputs it into the backbone network, enabling the network to dynamically adjust the weights of the first two convolutional kernels based on the scattering intensity of the current scene (such as sea fog, backlight, etc.). This preserves the edge and texture information of small targets in strong scattering scenarios and maintains normal semantic extraction efficiency in weak scattering scenarios, avoiding the feature loss problem caused by fixed convolutional kernels. The neck network outputs a high-resolution P2 feature map for detecting distant small targets. Combined with bidirectional feature fusion using learnable weights, it can adaptively adjust the fusion ratio of shallow details and deep semantics, reducing information loss caused by downsampling and improving the recall rate of targets with low pixel ratios. Through a dual-headed differential detection subnetwork in the uncertain band of the sea-line region, an uncertain band is dynamically constructed based on the estimated position of the sea-line region, temporal drift variance, and basic bandwidth. Independent detection branches are used for joint inference in the in-band and out-of-band regions, which can effectively distinguish false targets (such as wave crests and highlights) from real ships in the sea-line region, reducing the false detection rate while improving the detection stability of real targets. During the training phase, a label allocation strategy combining target scale and scattering intensity index is employed. This prioritizes positive sample matching for small targets in strong scattering scenarios, mitigating the problem of gradient dominance by medium and large targets. Simultaneously, a weighted joint loss function for small targets with center offset constraints is used to specifically suppress localization jitter for distant small targets, improving localization accuracy. The entire chain, from feature extraction and multi-scale fusion to differential detection, is collaboratively optimized for the scattering environment. It eliminates the need for external image enhancement preprocessing modules, avoiding noise or artifact interference introduced by enhancement algorithms, while maintaining the basic input-output paradigm of the YOLO framework, facilitating engineering deployment. Since the scattering intensity index is integrated into training, inference only requires inputting the image to be identified and its scattering intensity index. The network automatically performs environmental adaptive adjustment and joint inference, outputting stable bounding box coordinates and corresponding confidence scores for small target ships.

[0008] As a preferred technical means: the scattering intensity index mentioned in step 1) is obtained by weighted summation of multiple indices, and the sum of the weight coefficients of each index is 1; the multiple indices include the difference between 1 and the global contrast, the density of bright spots, the edge blurring and the fog coverage ratio.

[0009] This technical solution comprehensively quantifies the impact of the marine scattering environment from four dimensions: contrast reduction, speckle interference, edge blurring, and fog obstruction. It accurately reflects the combined effects of different scattering types, such as sea fog, salt spray, wave spray, and backlight diffusion. A unified scattering intensity index is constructed using a weighted summation method, with the sum of the weight coefficients being 1. This achieves normalized fusion of different scattering characteristics, providing a single, clear global prior signal for subsequent dynamic adjustment of the network. The complement of the fused global contrast directly reflects the overall degree of image sharpness reduction, complementing speckle density, edge blurring, and fog coverage ratio, thus avoiding the limitation of a single index in comprehensively characterizing scattering intensity.

[0010] As a preferred technical means, the step size of the P2 high-resolution feature map is 4.

[0011] By limiting the stride of the P2 feature map to 4, compared to the feature maps with strides of 8 or 16 in the conventional YOLO network, it has higher spatial resolution, can preserve the few pixels and edge details of small targets at a distance, and significantly improves the positive sample coverage and first detection probability of the small target detection head.

[0012] As a preferred technique, the specific method for dynamically adjusting the weights of convolutional kernels in the backbone network is as follows: multiple basic convolutional kernels are preset, and the dynamic combination weights of each basic convolutional kernel are calculated based on the input scattering intensity index. The multiple basic convolutional kernels are weighted and summed according to their corresponding weights to obtain the dynamic convolutional kernel for the current scene, and then feature extraction is performed through this dynamic convolutional kernel.

[0013] This technical solution uses a weighted combination of multiple basic convolutional kernels to enable the network to continuously adjust the feature extraction preference according to the scattering intensity (such as enhancing edge preservation when there is strong scattering and strengthening semantics when there is weak scattering). This is smoother and more flexible than simple convolutional kernel switching or channel attention, and does not significantly increase the number of parameters.

[0014] As a preferred technique, the dynamic combination weight is: after taking the exponential value of the gated response corresponding to each basic convolution kernel obtained by mapping the scattering intensity index, divide it by the sum of the exponential values ​​of all gated responses.

[0015] This technical solution uses a normalization method to calculate weights, ensuring that the sum of the weights of each basic convolution kernel is 1, and that the weight allocation is sensitive to changes in the scattering intensity index; the exponential operation amplifies the contribution of the high-response kernel, making the dynamic convolution kernel more biased towards the convolution mode most suitable for the current scene, thereby improving the effectiveness of feature extraction.

[0016] As a preferred technical means: in the bidirectional feature fusion neck network with learnable weights, the output feature map is equal to the sum of the corresponding learnable normalized weights multiplied by each input feature map after channel or scale alignment transformation; the learnable normalized weight is: the result of taking the exponential value of each learnable parameter, dividing it by the sum of the exponential values ​​of all learnable parameters, and adding it to a very small constant.

[0017] This technical solution uses learnable normalized weights, enabling the neck network to automatically adjust the fusion ratio of features at different scales (especially shallow high-resolution P2 features and deep semantic features) according to the data distribution, thus avoiding excessive masking of shallow details by deep semantics. Exponential normalization ensures the non-negativity of the weights and that their sum is 1, while the extremely small constant enhances numerical stability.

[0018] As a preferred technical means: In step 4), the P2 detection head includes two 3×3 convolutional layers with a stride of 1 and a padding of 1, and a classification and regression joint output layer. Its output features are upsampled and fused with the input features of the P3 detection head to achieve cross-scale enhancement of small target features. The parameters of the uncertain zone-specific detection branch and the main sea surface detection branch of the sea-line uncertain zone dual-head differential detection sub-network are initialized and trained independently. The uncertain zone-specific detection branch adds a center offset constraint sub-layer, and its classification layer threshold is set to 0.6 and the regression layer loss weight is set to 1.2. The classification layer threshold of the main sea surface detection branch is set to 0.5 and the regression layer loss weight is set to 1.0. During the inference process, the detection confidence threshold is adaptively adjusted according to the input scattering intensity index, wherein the larger the scattering intensity index, the lower the detection confidence threshold.

[0019] The P2 detection head in this technical solution employs a two-layer 3×3 convolutional layer design, which can fully extract local features of small targets from high-resolution feature maps. The joint classification and regression output layer simplifies the detection head structure and improves inference efficiency. Upsampling and fusion of the output features of the P2 detection head and the input features of the P3 detection head achieves complementarity between the detailed features of small targets and the semantic features of mesoscale targets, further enhancing the feature representation capability of small targets. The parameters of the dual-head detection branches are initialized and trained independently, enabling the two branches to learn the feature distribution of ordinary sea surface areas and uncertain sea-line areas respectively, avoiding mutual interference between detection requirements of different areas. The dedicated detection branch for uncertain zones uses a higher classification threshold and regression loss weights, and adds a center offset constraint sublayer, which can effectively suppress false target responses in the sea-line neighborhood while improving the positioning accuracy of small target ships. During inference, the confidence threshold is adaptively adjusted based on the scattering intensity. In strong scattering scenarios, the threshold is lowered to ensure the recall rate of small targets, while in weak scattering scenarios, the threshold is raised to suppress false detections, achieving a dynamic balance between recall and false detection rate.

[0020] As a preferred technical means: the specific method of the scale-aware label allocation strategy in step 3) is as follows: comprehensively consider the confidence of the predicted category, the overlap between the predicted box and the real box, whether the target area is less than the preset small target threshold, and the scattering intensity of the current scene to calculate the matching score of each candidate box. The dynamic Top-k mechanism is used to select the candidate box with the highest matching score as the positive sample, and the small target sample is given a higher center prior weight. The training samples are then sampled in buckets according to the scattering intensity.

[0021] This technical solution comprehensively considers four dimensions—category confidence, overlap, target scale, and scattering intensity—to calculate the matching score. This allows for a more comprehensive and accurate evaluation of the matching degree between candidate boxes and the real target, avoiding the problem of unreasonable positive sample allocation caused by single-dimensional evaluation. A dynamic Top-k mechanism is employed to select positive samples, automatically adjusting the number of positive samples based on the target size and scattering intensity, ensuring that small targets and strong scattering samples receive sufficient supervision signals. Small target samples are assigned higher center prior weights, making the network more inclined to select candidate boxes closer to the center of the real target as positive samples, thus improving the localization accuracy of small targets. Bucketing the training samples according to scattering intensity ensures that samples with different scattering intensities receive balanced learning during training, improving the model's generalization ability across various sea conditions.

[0022] As a preferred technical means: the weighted joint loss function for small targets including center offset constraints is the sum of multiple loss functions multiplied by their corresponding weight coefficients. The loss functions include classification loss, target confidence loss, bounding box regression loss, and center offset constraint loss. Each target sample is assigned a loss weight that is negatively correlated with the pixel area. The smaller the target pixel area, the greater the corresponding loss weight. The center offset constraint loss is used to suppress the jitter of the center position of the bounding box of small targets at a distance.

[0023] This technical solution employs a multi-loss function joint optimization approach, simultaneously considering classification accuracy, confidence reliability, bounding box regression accuracy, and center position stability, thus comprehensively improving detection performance. Assigning loss weights negatively correlated with pixel area to small target samples significantly increases their contribution to the total loss, effectively mitigating the problem of medium to large targets dominating gradient updates in conventional training. The center offset constraint loss specifically suppresses center position jitter in the bounding boxes of distant small targets, solving the industry-wide problem of unstable positioning of small targets on ships. The various loss functions are balanced through weight coefficients, allowing for flexible adjustment of the importance of different loss terms according to actual application needs, improving the model's customizability.

[0024] Secondly, a system for identifying small marine targets in a scattering environment based on YOLO-SOD is provided, comprising: The data acquisition and index calculation module is used to acquire marine images or video frames and calculate their scattering intensity index. The network construction module is used to build the YOLO-SOD network, which includes a scattering-aware dynamic convolutional backbone network, a bidirectional feature fusion neck network with learnable weights, a P2~P5 multi-scale detection head group, and a dual-head differential detection sub-network for the uncertain zone of the sea-line. The training module is used to input training samples and their scattering intensity indices into the YOLO-SOD network. It uses a scale-aware label allocation strategy that combines target scale and scattering intensity indices to perform positive sample matching, and uses a small target weighted joint loss function with center offset constraints to optimize the network parameters, thus obtaining the trained network. The inference and detection module is used to input the image to be identified and its scattering intensity index into the trained network, and execute the following steps in sequence: the backbone network dynamically adjusts the weights of the first two convolutional kernels based on the scattering intensity index and outputs multi-scale feature maps; the neck network performs adaptive weighted fusion of the feature maps and outputs P2 high-resolution feature maps and P3~P5 feature maps for detecting small targets at a distance; the detection head group detects each feature map to obtain preliminary features; the dual-head sub-network dynamically constructs an uncertain band based on the sea-line estimated position, temporal drift variance and basic bandwidth, and uses a dedicated branch within the band and a main sea surface branch outside the band to jointly infer the preliminary features to obtain the bounding box coordinates and confidence of the small target ship; The result output module is used to output the bounding box coordinates and confidence level.

[0025] The system uses the scattering intensity index as a global control signal throughout the entire process of data processing, model training, and inference detection. From a system architecture perspective, it achieves adaptability to different scattering sea conditions, effectively addressing the industry-wide common problem of high false positive and false negative rates for small targets at sea under strong scattering environments. The training module integrates a scale-aware label allocation strategy and a weighted joint loss function for small targets at sea, effectively strengthening the supervision signal for small target samples. The inference detection module integrates a dual-head differentiated detection architecture for the uncertain band of the sea-line, specifically optimized for the dense false target characteristics in the sea-line neighborhood, significantly reducing the false positive rate while maintaining overall recall.

[0026] Beneficial effects: 1. Improved recall rate for small-target ship detection at sea: Through the collaborative design of P2 small-target detection head and scattering-aware dynamic convolution, the shallow edges and weak texture information of low-pixel targets are effectively preserved, reducing the loss of details caused by downsampling and improving the network's response capability to low-contrast, small-sized ship targets.

[0027] 2. Optimization of detection accuracy in the neighborhood of the sea-line: Relying on the dual-head differential detection subnetwork of the uncertain band of the sea-line, differential modeling and regional fusion output are performed on the boundary region to reduce false target response, improve the detection stability of real small targets, and suppress false detection and missed detection problems in the neighborhood of the sea-line.

[0028] 3. Training supervision adapts to the needs of small target detection: Scale-aware label allocation, dynamic Top-k positive sample selection, and small target weighted loss enhance the training supervision contribution of small target samples, alleviate the problem of gradient dominance for medium and large targets, and fully release the feature expression capabilities of the network structure.

[0029] 4. Excellent engineering deployment and scalability: This method does not rely on image restoration networks, maintains the basic input-output paradigm of the YOLO framework, and can directly output the bounding box and confidence results of ship targets, which is convenient for integration with existing maritime intelligent sensing systems; at the same time, it can be extended to recognition systems, electronic devices, storage media and other implementation forms, which has engineering application value. Attached Figure Description

[0030] Figure 1 This is an overview diagram of the implementation method of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0032] Example 1 This embodiment provides a method for identifying small marine targets in a scattering environment based on YOLO-SOD, such as... Figure 1 As shown, the technical process of this invention can be summarized as follows: First, acquire marine images or video frames and complete basic preprocessing to calculate the scattering intensity index characterizing the scattering degree of the current scene; then, construct a YOLO-SOD network, which includes a scattering-aware dynamic convolutional backbone network, a bidirectional feature fusion neck network with learnable weights, P2~P5 multi-scale detection head groups, and a dual-head differential detection sub-network for the uncertain sea-line zone; input training samples and their scattering intensity index into the YOLO-SOD network, and train and optimize the network parameters using a scale-aware label allocation strategy and a small target weighted joint loss function; finally, in the inference stage, input the image to be identified and its scattering intensity index into the trained YOLO-SOD network, and output the bounding box coordinates and confidence results of the small target ship through scattering-aware feature extraction, multi-scale feature fusion, construction of the uncertain sea-line zone, and dual-head differential detection.

[0033] The specific steps are as follows: S1: Acquire images or video frames of the sea and calculate scattering intensity indices. This step aims to extract statistics from the input image that reflect the strength of the scattering environment, providing a unified prior for subsequent dynamic backbone adjustment, training sample binning, and inference threshold adaptation.

[0034] S1.1 Input Data Acquisition and Basic Preprocessing: Continuous maritime image frames are acquired from shore-based fixed cameras, port and waterway monitoring systems, or maritime patrol videos. The input images can be single frames or a sequence of frames sampled from a video stream at fixed time intervals. The raw input undergoes uniform size normalization, outlier frame removal, brightness range limitation, and color space normalization to reduce interference from overexposed, underexposed, or severely blurred frames in subsequent calculations.

[0035] Calculation of S1.2 scattering correlation statistics: Calculate the global contrast for each single frame image. High density of bright spots on the sea surface Edge blur With fog coverage ratio Among them, global contrast can be characterized by grayscale standard deviation or local contrast mean; bright spot density can be obtained by threshold segmentation statistics of high-brightness areas; blurriness can be characterized by the degree of gradient energy decay; fog coverage ratio can be estimated by the proportion of low saturation and low contrast areas.

[0036] Construction of S1.3 scattering intensity index: The above statistics are combined into a unified scattering intensity index.

[0037] in, For normalized weight coefficients, satisfying The higher the value of this indicator, the stronger the influence of sea fog, salt spray, wave spray, and backlight diffusion on the current scene. Subsequent scattering-related modules in the network structure can all use this indicator as an adjustment parameter.

[0038] S2: Constructing the YOLO-SOD network This step is the core structural implementation stage. The goal is to improve the detection capability and positioning stability of small target vessels at sea by constructing a YOLO-SOD network structure, without relying on image restoration networks.

[0039] S2.1 Scatter-sensing dynamic backbone construction: A scatter-aware dynamic convolution module is introduced into the first and second convolutional layers of the YOLO backbone network. This module adaptively adjusts the convolutional kernel combination based on the scattering intensity index Is, instead of using fixed convolutional kernels. This enhances the edge preservation capability in strong scattering scenarios and maintains semantic extraction efficiency in weak scattering scenarios. The expression is as follows:

[0040] in,

[0041] In the formula, To preset the total number of basic convolutional kernels, Indicates the first One basic convolutional kernel, The gated response is obtained by mapping from the scattering intensity index. For the first The combined weights of the basic convolutional kernels. This design mitigates the degradation of weak textures on small targets by conventional backbones during early downsampling.

[0042] S2.2 Learnable Weights Bidirectional Feature Fusion Neck Construction: A bidirectional feature fusion structure with learnable weights is employed in the neck region to bidirectionally transfer shallow high-resolution features and deep strong semantic features. The fusion relationship is as follows:

[0043] in,

[0044] In the formula, N is the number of feature maps participating in the fusion. The input is a multi-scale feature map. This is a channel and scale alignment transformation. For learnable normalized weights, For learnable parameters, It is a very small positive number. This structure can automatically adjust the fusion ratio of features at different scales according to the data distribution, so that the shallow details related to small targets are not overly masked by deep smooth semantics during multi-scale propagation.

[0045] S2.3 P2 Small Target Detection Head Settings: A new P2 small target detection head with a stride of 4 is added to the detection layer, enabling the network to directly utilize higher-resolution feature maps to detect distant ship targets with a low pixel ratio. The P2 detection head, together with the P3, P4, and P5 detection heads, constitutes a multi-scale detection system. The P2 layer is primarily responsible for detecting ultra-small and small targets, while the P3-P5 layers are mainly responsible for detecting medium-scale targets. This setup effectively improves the positive sample coverage and first-detection probability of distant small targets.

[0046] Construction of the dual-head differential detection subnetwork for S2.4 sea-line uncertainty band: A main sea surface detection branch and an uncertain zone-specific detection branch are set up at the detection end. The main sea surface detection branch is used to process candidate targets in the ordinary sea surface area, while the uncertain zone-specific detection branch is used to process candidate targets in the uncertain zone neighborhood of the sea-line. The two branches share the front-end backbone network and some neck fusion features, but the detection output layer parameters are independent of each other. The uncertain zone-specific detection branch adds a center offset constraint sublayer to enhance the positioning stability of small targets in the sea-line neighborhood and suppress interference from false targets such as wave crests, highlights, and cloud edges.

[0047] S3: Training the network (based on scale-aware label assignment and small target weighted loss) S3.1 Scale-Aware Label Allocation: The matching score for each candidate box is calculated by combining the confidence score of the predicted category, the overlap between the predicted and ground truth boxes, whether the target area is smaller than a preset small target threshold, and the scattering intensity of the current scene.

[0048] in, For the first The candidate box corresponds to the first The confidence score of the predicted category of each real target. For the first The candidate box and the first The intersection-union ratio of a real bounding box. For the first The pixel area of ​​a real target. For small target area threshold, For category confidence weight coefficients, For the intersection-union ratio weighting coefficient, For the small target weight coefficient, This is the scattering intensity weighting coefficient. The indicator function is used. A dynamic Top-k mechanism is employed to select the candidate box with the highest matching score as the positive sample, and higher center prior weights are assigned to small target samples. The training samples are then binned according to scattering intensity.

[0049] S3.2 Small Objective Weighted Joint Loss Function: The training loss includes at least classification loss, object confidence loss, bounding box regression loss, and center offset constraint loss, and their joint expression is:

[0050] in, For classifying losses, Confidence loss for the target For bounding box regression loss, For center offset constraint loss, Classification loss weights, Weight the target confidence loss. For bounding box regression loss weights, The center offset constraint loss weights are used. To enhance the supervision strength of small targets, the weights are adjusted for the ... Each objective introduces area-related weights:

[0051] in, This is the area normalization coefficient. This is the weighting amplification factor. The smaller the target area, the greater its loss weight.

[0052] S3.3 training implementation: During the training phase, high-resolution input sizes such as 960×960 or 1280×1280 can be used, with higher input resolution preferred for scenarios where small objects account for a large proportion. The optimizer can be SGD or AdamW, and the learning rate can employ a cosine annealing or piecewise decay strategy.

[0053] S4: Inference Detection This step utilizes the trained YOLO-SOD network to detect small target vessels in the maritime image or video frame to be identified. During the inference phase, the image or video frame to be identified is first acquired, and the scattering intensity index of the current scene is calculated as described in S1. Then, the image to be identified and its scattering intensity index are input into the trained YOLO-SOD network. The scattering-aware dynamic convolutional backbone network performs multi-scale feature extraction, the bidirectional feature fusion neck network with learnable weights outputs the P2-P5 fused feature map, and the multi-scale detection head group generates preliminary candidate detection results. Finally, combined with the dual-head differential detection sub-network for the uncertain sea-line region, regional joint inference is performed on the candidate targets, outputting the bounding box coordinates and confidence score of the small target vessel.

[0054] S4.1 Candidate Detection Results Generation: The image to be identified is denoted as The corresponding scattering intensity is labeled as .Will and Input the trained YOLO-SOD network, and the backbone network according to... The weights of the first two convolutional kernels are dynamically adjusted, and multi-scale backbone features are output. The neck network performs learnable weighted fusion of the multi-scale features, outputting P2, P3, P4, and P5 fused feature maps. The detection head group performs detection on the above fused feature maps to obtain a preliminary set of candidate detection results.

[0055] in, Indicates the first One candidate test result, Indicates the first The coordinates of each candidate box. This indicates the corresponding confidence level. Indicates the corresponding category, This indicates the initial number of candidate boxes.

[0056] S4.2 Sea Line Uncertain Band Structure: The sea-line position is estimated from the input image, and the center curve of the sea-line is obtained. In continuous frame scenarios, the temporal drift variance of the sea-line position is statistically analyzed. Constructing the uncertain band region of the sea-line:

[0057] in, Based on bandwidth, This is the variance amplification factor. This represents the temporal drift variance of the sea-line position. By defining this region, the most error-prone strip area near the sea-line can be explicitly extracted.

[0058] S4.3 Regionalized Fusion Output: The final test results adopted a regional fusion strategy:

[0059] in, For uncertain output from the detection head, Output from the main sea surface detection head. This is the indicator function. Candidate boxes located within the uncertainty zone are preferentially selected from the uncertainty zone detector output, while candidate boxes located in other areas are preferentially selected from the main sea surface detector output. This reduces false detections in the sea-line neighborhood caused by wave crests, highlights, and cloud edges, while increasing the probability of retaining real small vessel targets.

[0060] S4.4 Adaptive Confidence Threshold and Result Output: During the inference phase, the detection confidence threshold is adaptively adjusted based on the scattering intensity index.

[0061] in, Based on the confidence threshold, is the adjustment coefficient. The threshold is appropriately lowered when scattering is enhanced to ensure recall, and the threshold is increased when scattering is weakened to suppress redundant candidate boxes. The final recognition result can be expressed as:

[0062] in, Indicates the first One detection box, This indicates the corresponding confidence level. This indicates the number of valid targets detected in the current frame.

[0063] To verify the effectiveness of each module in this invention, multiple ablation experiments were conducted using the standard YOLO model as a baseline for comparison. These experiments verified the effects of adding only the P2 detector head, adding a dynamic adjustment strategy for the scattering intensity index, adding a detector head for the uncertain zone of the sea-line, and the complete YOLO-SOD model integrating the scattering sensing backbone, dual-head detection of the sea-line, and small target weighted loss. Through ablation comparison, the specific contributions of each module to the detectability of distant small targets (ships) in scattering environments, the robustness of the sea-line neighborhood, and the positioning stability can be verified. The data are shown in Table 1. Table 1 Qualitative Comparison Table

[0064] As shown in Table 1, the overall ship detection performance shows an improving trend with the gradual improvement of the model structure. The standard YOLO model, as the baseline model, performs relatively poorly in terms of recall and mAP@0.5 (mean precision, IoU threshold 0.5). After adding the P2 detector head, the dynamic adjustment strategy for the scattering index, and the sea-sky uncertainty zone detector head, the model's detection performance is improved to varying degrees. The complete YOLO-SOD model, after integrating the above improvements and adding a small target weighted loss, achieves the highest results in recall and mAP@0.5, significantly improving the accuracy and stability of identifying small targets at long distances in scattering environments.

[0065] Without departing from the core idea of ​​this invention, the above modules can also be implemented using equivalent substitutions. For example, the scattering-sensing dynamic convolution can be replaced by conditional convolution, expert hybrid convolution, or a channel recalibration module; the weighted bidirectional feature fusion structure can be replaced by other feature pyramid structures with learnable fusion weights; the sea-line uncertainty band can be provided by geometric estimation, lightweight segmentation networks, or temporal boundary estimation modules. None of the above substitutions change the core concept of "improving the ability to identify small, long-range targets at sea in scattering environments by improving the YOLO network structure," and should be considered within the scope of protection of this invention.

[0066] This invention provides a complete, feasible, reproducible, and deployable technical solution for identifying small targets at sea under scattering sea conditions through a joint design of "scattering perception backbone modification + weighted bidirectional neck fusion + detection end differentiated modeling + training supervision rebalancing". This solution is compatible with the YOLO engineering ecosystem and is easy to implement in systems such as maritime supervision and coastal defense perception.

[0067] Example 2 This embodiment provides a YOLO-SOD-based system for identifying small marine targets in scattering environments, comprising: The data acquisition and index calculation module is used to acquire marine images or video frames and calculate scattering intensity indices. The network construction module is used to construct the YOLO-SOD network, which includes a scattering-aware dynamic convolutional backbone network, a bidirectional feature fusion neck network with learnable weights, a P2~P5 multi-scale detection head group, and a dual-head differential detection sub-network for uncertain bands in the sea-line region. The training optimization module is used to input training samples and their scattering intensity indices into the YOLO-SOD network, and optimize the network parameters using a scale-aware label allocation strategy and a small target weighted joint loss function to obtain the trained YOLO-SOD network. The inference detection module is used to input the image to be identified and its scattering intensity index into the trained YOLO-SOD network, and sequentially complete the extraction of scattering sensing features, multi-scale feature fusion, construction of the sea-line uncertainty band, and dual-head differential detection. The results output module is used to output the bounding box coordinates and corresponding confidence scores of small target ships.

[0068] It is understood that the detailed functional implementation of the above modules can be found in the description of the aforementioned method embodiments, and will not be elaborated further here.

[0069] The above are specific embodiments of the present invention, which demonstrate the substantial features and progress of the present invention. Based on the actual needs of use, equivalent modifications in shape, structure, etc., can be made to it according to the teachings of the present invention, and all such modifications are within the scope of protection of this solution.

Claims

1. A method for identifying small maritime targets based on YOLO-SOD scattering environment, characterized in that, The steps include the following: 1) Acquire images or video frames at sea and calculate their scattering intensity index; 2) Construct the YOLO-SOD network, which includes a scattering-sensing dynamic convolutional backbone network, a bidirectional feature fusion neck network with learnable weights, a P2~P5 multi-scale detection head group, and a dual-head differential detection sub-network for uncertain sea-line areas; 3) Train the YOLO-SOD network by inputting training samples and their scattering intensity indices into the YOLO-SOD network. A scale-aware label allocation strategy combining target scale and scattering intensity indices is used for positive sample matching. The network parameters are optimized using a small target weighted joint loss function with center offset constraints. The specific method of the scale-aware label allocation strategy is as follows: the matching score of each candidate box is calculated by comprehensively considering the prediction class confidence, the overlap between the predicted box and the ground truth box, whether the target area is less than a preset small target threshold, and the current scene scattering intensity. A dynamic Top-k mechanism is used to select the candidate box with the highest matching score as a positive sample. Small target samples are given higher center prior weights, and the training samples are sampled by bucketing according to scattering intensity. 4) Input the image to be identified and its scattering intensity index into the trained network: The backbone network dynamically adjusts the weights of its first two convolutional kernels based on the scattering intensity index and outputs multi-scale feature maps; the neck network performs adaptive weighted fusion of the feature maps and outputs P2 high-resolution feature maps and P3~P5 feature maps for detecting small targets at a distance; the detection head group detects each feature map to obtain preliminary features; the sea-line uncertain band dual-head differential detection sub-network dynamically constructs uncertain bands based on the sea-line estimated position, temporal drift variance and basic bandwidth, and uses a dedicated branch within the uncertain band and a main sea surface branch outside the band to jointly infer the preliminary features, outputting the bounding box coordinates and confidence of the small target ship; The P2 detection head consists of two 3×3 convolutional layers with a stride of 1 and padding of 1, and a joint classification and regression output layer. Its output features are upsampled and fused with the input features of the P3 detection head to achieve cross-scale enhancement of small target features.

2. The method for identifying small maritime targets in a scattering environment based on YOLO-SOD according to claim 1, characterized in that: The scattering intensity index mentioned in step 1) is obtained by weighted summation of multiple indices, and the sum of the weight coefficients of each index is 1; the multiple indices include the difference between 1 and the global contrast, the density of bright spots, the edge blurring and the fog coverage ratio.

3. The method for identifying small maritime targets in a scattering environment based on YOLO-SOD according to claim 1, characterized in that: The step size of the P2 high-resolution feature map is 4.

4. The method for identifying small maritime targets in a scattering environment based on YOLO-SOD according to claim 1, characterized in that: The specific method for dynamically adjusting the weights of convolutional kernels in the backbone network is as follows: multiple basic convolutional kernels are preset, and the dynamic combination weights of each basic convolutional kernel are calculated based on the input scattering intensity index. The multiple basic convolutional kernels are weighted and summed according to their corresponding weights to obtain the dynamic convolutional kernel for the current scene, and then feature extraction is performed through this dynamic convolutional kernel.

5. The method for identifying small maritime targets in a scattering environment based on YOLO-SOD according to claim 4, characterized in that: The dynamic combination weight is: after taking the exponential value of the gated response corresponding to each basic convolution kernel obtained by mapping the scattering intensity index, divide it by the sum of the exponential values ​​of all gated responses.

6. The method for identifying small maritime targets in a scattering environment based on YOLO-SOD according to claim 1, characterized in that: In the bidirectional feature fusion neck network with learnable weights, the output feature map is equal to the sum of the corresponding learnable normalized weights multiplied by each input feature map after channel or scale alignment transformation; the learnable normalized weights are: the result of taking the exponential value of each learnable parameter, dividing it by the sum of the exponential values ​​of all learnable parameters, and adding it to a very small constant.

7. The method for identifying small maritime targets in a scattering environment based on YOLO-SOD according to claim 1, characterized in that: In step 4), the parameters of the uncertain zone-specific detection branch and the main sea surface detection branch of the uncertain zone dual-head differential detection subnetwork are initialized and trained independently. The uncertain zone-specific detection branch adds a center offset constraint sublayer, with its classification layer threshold set to 0.6 and regression layer loss weight set to 1.

2. The classification layer threshold of the main sea surface detection branch is set to 0.5 and regression layer loss weight set to 1.

0. During the inference process, the detection confidence threshold is adaptively adjusted according to the input scattering intensity index, wherein the larger the scattering intensity index, the lower the detection confidence threshold.

8. The method for identifying small maritime targets in a scattering environment based on YOLO-SOD according to claim 1, characterized in that: The weighted joint loss function for small targets with center offset constraints is the sum of multiple loss functions multiplied by their corresponding weight coefficients. The loss functions include classification loss, target confidence loss, bounding box regression loss, and center offset constraint loss. Each target sample is assigned a loss weight that is negatively correlated with its pixel area. The smaller the target pixel area, the greater the corresponding loss weight. The center offset constraint loss is used to suppress the jitter of the center position of the bounding box of small targets at a distance.

9. A system for identifying small marine targets in a scattering environment based on YOLO-SOD, characterized in that, The system employs the YOLO-SOD-based method for identifying small marine targets in a scattering environment as described in any one of claims 1-8, the system comprising: The data acquisition and index calculation module is used to acquire marine images or video frames and calculate their scattering intensity index. The network construction module is used to build the YOLO-SOD network, which includes a scattering-aware dynamic convolutional backbone network, a bidirectional feature fusion neck network with learnable weights, a P2~P5 multi-scale detection head group, and a dual-head differential detection sub-network for the uncertain zone of the sea-line. The training module is used to input training samples and their scattering intensity indices into the YOLO-SOD network. It uses a scale-aware label allocation strategy that combines target scale and scattering intensity indices to perform positive sample matching, and uses a small target weighted joint loss function with center offset constraints to optimize the network parameters, thus obtaining the trained network. The inference and detection module is used to input the image to be identified and its scattering intensity index into the trained network, and execute the following steps in sequence: the backbone network dynamically adjusts the weights of the first two convolutional kernels based on the scattering intensity index and outputs multi-scale feature maps; the neck network performs adaptive weighted fusion of the feature maps and outputs P2 high-resolution feature maps and P3~P5 feature maps for detecting small targets at a distance; the detection head group detects each feature map to obtain preliminary features; the dual-head sub-network dynamically constructs an uncertain band based on the sea-line estimated position, temporal drift variance and basic bandwidth, and uses a dedicated branch within the band and a main sea surface branch outside the band to jointly infer the preliminary features to obtain the bounding box coordinates and confidence of the small target ship; The result output module is used to output the bounding box coordinates and confidence level.

Images