A feature enhancement method, system and related equipment for marine target detection
By introducing a difference prediction model to guide convolution kernel sampling, the problem of insufficient feature extraction in marine target detection under complex sea conditions is solved, and high-precision and stable target detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YANTAI UNIV
- Filing Date
- 2025-12-24
- Publication Date
- 2026-08-04
AI Technical Summary
In complex sea conditions, traditional convolutional neural networks based on fixed sampling grids are insufficient in feature extraction for maritime target detection, which easily leads to missed detections and false detections. Furthermore, existing deformable convolutional methods lack clear physical constraints, causing confusion between target and background information during feature fusion.
We introduce interpretable, data-driven physical guidance signals and use a difference prediction model to predict the degree of inter-class difference between the target and the background. This prediction serves as a criterion for selecting sampling points for deformable convolution kernels, avoiding feature mutation boundary regions, and focusing on homogeneous internal regions for feature aggregation.
It improves feature quality, significantly increases recall and accuracy in maritime target detection, and reduces false detection rate, especially performing well in the detection of small targets with low contrast and irregular shapes.
Smart Images
Figure CN121391645B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision, image processing and remote sensing technology. Specifically, it relates to a feature enhancement method, a corresponding system and related computing equipment and media for marine target detection in complex sea conditions. It is particularly suitable for high-precision and robust automatic identification of targets such as ships, sea ice and islands. Background Technology
[0002] Maritime target detection is a core technology for ensuring maritime safety, conducting marine monitoring, and resource development. With the widespread adoption of satellite remote sensing and UAV aerial photography, the amount of available marine image data has surged, creating an urgent need for automated and intelligent detection systems.
[0003] However, the complex and ever-changing marine environment presents inherent challenges to target detection: First, the contrast between the target (such as a small vessel) and the background (sea surface) is low, and the boundaries are blurred. Second, interference from waves, ripples, cloud shadows, etc., creates a large amount of target-like noise. Third, the target itself has irregular shapes and varying scales. Traditional convolutional neural networks based on fixed sampling grids are insufficient in feature extraction in such scenarios, easily leading to missed detections and false positives.
[0004] Deformable convolutions enhance a model's adaptability to geometric deformations by introducing learnable offsets. However, most existing methods directly regress the offsets, lacking explicit physical constraints. In complex texture backgrounds, they are prone to learning suboptimal or even incorrect sampling points, leading to confusion between target and background information during feature fusion and limiting further performance improvements.
[0005] Therefore, the industry urgently needs an enhancement technology that can intelligently perceive image content and provide reliable prior guidance for feature sampling in order to improve the representation capability of features and thus achieve stable and accurate detection of maritime targets in complex sea conditions. Summary of the Invention
[0006] This invention aims to overcome the aforementioned deficiencies of existing technologies and provide a feature enhancement method, system, and related equipment for maritime target detection. The core technology of this invention is to introduce an interpretable, data-driven physical guidance signal into the feature sampling process. A specially trained model predicts the average difference between the target and the background in the image (i.e., the predicted inter-class difference value), and uses this value as a criterion for determining whether local sampling points are homogeneous with the center point. This guides the selection of sampling points for deformable convolutional kernels, enabling them to actively avoid boundary regions with abrupt feature changes and focus on homogeneous internal regions for feature aggregation, thereby improving feature quality from the source.
[0007] To achieve the above objectives, the present invention provides the following technical solution: A feature enhancement method for maritime target detection, the feature enhancement method comprising the following steps: Step S1 Image Processing: Acquire remote sensing images containing maritime targets and perform normalization preprocessing on the images; Step S2 Feature Extraction: Input the preprocessed image into a convolutional neural network for feature extraction to obtain a multi-channel feature map; Step S3: Difference prediction: Input the multi-channel feature map into a pre-trained difference prediction model, which outputs an inter-class difference prediction value to characterize the degree of difference between the target and the background. Step S4 Adaptive Sampling: For each candidate sampling point of the deformable convolution kernel, calculate the inter-point difference value between it and the center point; using the predicted inter-class difference value as a criterion, select candidate sampling points whose inter-point difference value is less than the predicted inter-class difference value to form a homogeneous sampling point set; if the number of homogeneous sampling points does not meet the requirements of the convolution operation, supplement the remaining candidate sampling points with the smallest inter-point difference value from the remaining candidate sampling points whose inter-point difference value is greater than or equal to the predicted inter-class difference value to form the final sampling point set for the convolution operation; perform deformable convolution operation based on the final sampling point set to extract features; Step S5 Target Detection: Fuse features extracted based on the final set of sampling points with features extracted based on standard grid sampling points, input the features into the target detection network, and perform maritime target detection; The maritime targets include at least one of sea ice, ships, and islands; the remote sensing images are derived from at least one of satellite remote sensing, UAV aerial photography, or shipborne radar imaging.
[0008] Preferably, the difference prediction model in step S3 makes predictions by fusing the local multidimensional statistical features of the multi-channel feature map with the global contextual features; the local multidimensional statistical features include: standard deviation, mean, contrast, entropy, spatial frequency, and fractal dimension.
[0009] Preferably, the difference prediction model includes: The local feature extraction branch is used to process the local multidimensional statistical features through a multilayer perceptron to generate local descriptive factors; The global feature extraction branch is used to process the multi-channel feature map through the Transformer module to generate a global descriptor. The feature fusion and regression module is used to fuse the local descriptive factors and the global descriptive factors, and to regress and output the predicted inter-class differences.
[0010] Preferably, the difference prediction model is obtained through the following supervised training method: Obtain a training set of remote sensing images with pixel-level annotations and extract their multi-channel feature maps; For the multi-channel feature map, the target region and background region on the feature map are mapped according to the annotation of the corresponding original image. The feature mean of the two regions on the feature map is calculated respectively, and the norm between the two feature means is calculated as the actual inter-class difference ground truth label of the feature map.
[0011] Using the multi-channel feature map as input and the actual inter-class difference ground truth labels as supervision targets, the mean squared error loss function is used to train the model end-to-end.
[0012] Preferably, step S4 includes the following steps: S41: Obtain the initial candidate sampling point positions of the deformable convolution kernel; S42: Calculate the difference between each candidate sampling point and the center point; S43: Compare the inter-point difference value with the inter-class difference prediction value, and filter out candidate sampling points whose inter-point difference value is less than the inter-class difference prediction value to form a homogeneous sampling point set; S44: Determine whether the number of sampling points in the homogeneous sampling point set meets the preset convolution kernel size requirement; if not, sort the remaining candidate sampling points from the remaining candidate sampling points whose inter-point difference value is greater than or equal to the inter-class difference prediction value according to the size of their inter-point difference value, and add the candidate sampling points with the smallest inter-point difference value to the homogeneous sampling point set in turn until the quantity requirement is met, forming the final sampling point set. S45: Perform deformable convolution operation based on the final set of sampling points.
[0013] Preferably, step S5 includes the following steps: S51: Based on the offset of the final set of sampling points relative to the standard grid, the feature map is divided into edge regions and non-edge regions; S52: Convolve the edge region using the final set of sampling points to extract edge features; S53: Convolution is performed on non-edge regions using standard grid sampling points to extract internal texture features; S54: The edge features and the internal texture features are fused to generate an enhanced feature map, which is then used by the target detection network to obtain the detection result.
[0014] The present invention also provides a feature enhancement system for maritime target detection, characterized in that it comprises: The image processing module is used to acquire and preprocess remote sensing images containing maritime targets; The feature extraction module, connected to the image processing module, is used to extract multi-channel feature maps from the preprocessed image; The difference prediction module is connected to the feature extraction module and has a pre-trained difference prediction model embedded in it. It is used to receive the multi-channel feature map and output the inter-class difference prediction value. An adaptive sampling module is connected to the feature extraction module and the difference prediction module respectively, and is used to filter sampling points according to the inter-class difference prediction value to form a final sampling point set. The target detection module, connected to the adaptive sampling module, is used to extract and fuse features based on the final set of sampling points to perform maritime target detection.
[0015] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described feature enhancement method for maritime target detection.
[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described feature enhancement method for maritime target detection.
[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described feature enhancement method for maritime target detection.
[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. Clear and reliable guiding logic: By predicting the inter-class difference value and directly using it as the homogeneity criterion, it provides a decision basis with clear physical meaning and interpretability for adaptive sampling, forming a logical closed loop from global semantic understanding to local operation execution.
[0019] 2. Fundamental improvement in feature quality: The adaptive sampling mechanism forces convolution operations to be performed only in regions that meet the homogeneity condition, which greatly avoids invalid sampling across the target-background boundary, resulting in higher feature purity and less noise.
[0020] 3. Enhanced edge preservation capability: By distinguishing between edge and non-edge regions and employing a differentiated sampling strategy, the target contour can be captured more precisely, mitigating the negative impact of boundary blurring.
[0021] 4. Significantly improved detection performance: Especially for small targets with low contrast and irregular shapes, and in harsh sea conditions, this invention can steadily improve the detection recall and accuracy, reduce false alarms, and has extremely high practical value.
[0022] Other features and advantages of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments or related technologies of this disclosure, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0024] Figure 1 This is an overall flowchart of the feature enhancement method for maritime target detection according to the present invention.
[0025] Figure 2 This is a schematic diagram illustrating the training process and application principle of the difference prediction model of this invention; Figure 3 This is a detailed logic flowchart of the adaptive sampling step of the present invention; Figure 4 This is a detailed flowchart of the feature fusion and target detection steps of the present invention; Figure 5 This is a structural diagram of the maritime target feature enhancement system module of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.
[0028] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatial relative descriptions used herein will be interpreted accordingly.
[0029] Please see Figure 1-5 The present invention provides a technical solution: like Figure 1 As shown, a feature enhancement method for maritime target detection includes the following steps: Step S1 Image Processing: Acquire a remote sensing image containing maritime targets and perform standardized preprocessing on the image; the maritime targets include at least one of sea ice, ships, and islands; the remote sensing image originates from at least one of satellite remote sensing, UAV aerial photography, or shipborne radar imaging. Optionally, the acquired image may be preprocessed, including but not limited to: image size normalization (e.g., uniform scaling to 512×512 pixels), color space conversion (e.g., converting to grayscale or retaining RGB), histogram equalization to enhance contrast, and using filters (e.g., Gaussian filtering) to suppress random noise. Preprocessing aims to provide standardized, high-quality input for subsequent steps.
[0030] Step S2 Feature Extraction: Input the preprocessed image into a pre-trained deep convolutional neural network (e.g., using ResNet-50 or VGG16 as the backbone network) for feature extraction. Remove the top-level classifier of the network and use its convolutional layers to extract rich multi-level spatial semantic features, outputting a multi-channel feature map F with dimensions [C, H, W], where C is the number of channels (e.g., 2048), and H and W are the height and width of the feature map, respectively.
[0031] Step S3: Difference prediction: Input the multi-channel feature map F into the pre-trained difference prediction model, which outputs a prediction value, namely the inter-class difference prediction value T. The physical meaning of the inter-class difference prediction value T: This value represents the average difference intensity of the feature responses between the foreground target and the background sea surface, predicted by the model after analyzing the entire feature map F. The larger the value, the higher the distinction between the target and the background in the current image.
[0032] Training of the differential prediction model (see) Figure 2 The training of this model is a supervised learning process.
[0033] 1. Data and Labels: A large number of precisely labeled remote sensing images are used. A dataset of remote sensing images with pixel-level annotations is employed. For each training image, after extracting its feature map, the mean features of the target region and the background region in the image are calculated based on the pixel-level annotation information of the corresponding original image. The norm of the difference between these two feature means is calculated, which is the "true value of the actual inter-class difference" for that image.
[0034] 2. Model Structure: The model employs a dual-path architecture. One path is a local statistical path, calculating six high-dimensional statistics of F: standard deviation, mean, contrast, entropy, spatial frequency, and fractal dimension, encoded by a multilayer perceptron. The other path is a global context path, using a lightweight Transformer module to capture long-range spatial dependencies in the feature maps. After the features from both paths are fused, a prediction threshold T_pred is output through a fully connected regression layer.
[0035] 3. Training and Optimization: Aiming to approximate the true label with the predicted value T_pred, the mean squared error loss function is used, and end-to-end training is performed using the Adam optimizer. After training, the model has the ability to accurately evaluate the uniformity of any local region based on the input feature map.
[0036] Step S4 Adaptive sampling (see...) Figure 3 This step uses the predicted inter-class difference prediction value T as a homogeneity criterion to intelligently select convolution sampling points.
[0037] * S41: For a standard 3×3 convolutional kernel with 9 regular grid locations, an initial offset field is predicted through an additional lightweight convolutional layer, generating a set of candidate sampling points {P_candidate} for each location (except the center).
[0038] * S42: For each candidate point p_i∈{P_candidate}, calculate the norm of the difference between the feature vector at the corresponding position on the feature map F and the feature vector at the center point, denoted as the difference value d_i between points.
[0039] * S43: Compare each d_i with the inter-class difference prediction value T. Retain all candidate sampling points that satisfy d_i < T. These points are considered to be "sufficiently similar" to the center point and are within a homogeneous region. They form the set of homogeneous sampling points {P_homogeneous}.
[0040] * S44: Check the number of points in {P_homogeneous}. If the number is equal to the number of neighborhood points required by the convolution kernel (e.g., 8), directly use it as the final set of sampling points {P_final}. If the number is insufficient, select the candidate sampling points with the smallest inter-point difference values from the eliminated candidate points for supplementation to ensure that the number of {P_final} is fixed and meets the requirements of convolution operations.
[0041] * S45: Use the positions indicated by {P_final} to perform deformable convolution operations on the feature map F, generating an adaptively enhanced primary feature response.
[0042] Step S5 Object detection (see Figure 4 ): This step completes the fusion of features and the final detection.
[0043] * S51: According to the offset magnitudes of each sampling point in {P_final} relative to the original regular grid, each position on the feature map can be automatically classified as an edge region (significant offset) or a non-edge region (tiny or zero offset).
[0044] * S52: For positions identified as edge regions, use {P_final} for convolution operations to focus on extracting the edge feature map F_edge that depicts the target contour.
[0045] * S53: For non-edge regions, use standard 3×3 regular grid sampling points for convolution to stably extract the texture feature map F_texture of the target interior or background region.
[0046] * S54: Concatenate or weighted fuse F_edge and F_texture in the channel dimension to form the final enhanced feature map F_enhanced. This feature map contains both clear boundary information and rich internal information. Input F_enhanced into the detection head of any modern object detection network (such as Faster R-CNN, RetinaNet, or the YOLO series) to output the bounding box coordinates of each maritime target and its所属类别 (such as ships, sea ice).
[0047] System embodiment: As Figure 5 shown, the present invention also provides a feature enhancement system 500 for maritime target detection. The system 500 includes: The image processing module is used to perform step S1.
[0048] The feature extraction module is connected to the image processing module and is used to perform step S2.
[0049] The difference prediction module is connected to the feature extraction module and is used to perform step S3.
[0050] The adaptive sampling module is connected to the feature extraction module and the difference prediction module respectively, and is used to execute step S4.
[0051] The target detection module is connected to the adaptive sampling module and is used to execute step S5.
[0052] Each module can be implemented through software, hardware, or a combination of both, working together to complete the entire process from raw image input to target detection result output.
[0053] Device, medium, and product embodiments: A computer device includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the aforementioned method steps. A computer-readable storage medium (such as an SSD, USB flash drive, or optical disc) stores a computer program implementing the aforementioned method. A computer program product comprising the computer program. These embodiments enable the invention to be easily deployed in servers, edge computing devices, or embedded systems.
[0054] Application Scenario Verification Example: This invention was deployed on a server cluster in a maritime monitoring center. The system processes video streams from multiple sensors used for maritime monitoring in real time. Under complex sea conditions such as low light, compared with traditional methods based on standard convolution or deformable convolution, the method of this invention can significantly improve the recall rate for detecting targets such as small vessels and effectively reduce the false detection rate. Experiments show that this invention improves feature quality from the source by guiding homogeneous region sampling through inter-class difference prediction values, thereby achieving more stable and accurate detection of maritime targets.
[0055] This invention offers a complete technical solution and a clear implementation route, and can be implemented on existing hardware platforms. Extensive experimental verification has demonstrated excellent performance and stability in maritime target detection missions, and it can be widely applied to: 1. Marine environmental monitoring and protection; 2. Automatic Identification and Tracking System for Ships; 3. Maritime search and rescue and emergency response; 4. Marine resource exploration and development; 5. Intelligent ship navigation and obstacle avoidance.
[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A feature enhancement method for maritime target detection, characterized in that, Includes the following steps: Step S1 Image Processing: Acquire remote sensing images containing maritime targets and perform normalization preprocessing on the images; Step S2 Feature Extraction: Input the preprocessed image into a convolutional neural network for feature extraction to obtain a multi-channel feature map; Step S3: Difference prediction: The multi-channel feature map is input into a pre-trained difference prediction model, which outputs an inter-class difference prediction value, which represents the average degree of difference between the target region and the background region in the feature map. Step S4 Adaptive Sampling: For each candidate sampling point of the deformable convolution kernel, calculate the inter-point difference value between it and the center point; filter out all candidate sampling points whose inter-point difference value is less than the predicted inter-class difference value to form a homogeneous sampling point set; if the number of the homogeneous sampling point set does not meet the requirements of the convolution operation, supplement the remaining candidate sampling points whose inter-point difference value is greater than or equal to the predicted inter-class difference value with the candidate sampling point with the smallest inter-point difference value to form the final sampling point set for the convolution operation; perform deformable convolution operation based on the final sampling point set to extract features; Step S5 Target Detection: Fuse the features extracted based on the final set of sampling points with the features extracted based on the standard grid sampling points, input the features into the target detection network, and output the location and category information of the maritime target; The maritime targets include at least one of ships, sea ice, and islands; The difference prediction model in step S3 makes predictions by fusing the local multidimensional statistical features and global contextual features of the multi-channel feature map; the local multidimensional statistical features include: standard deviation, mean, contrast, entropy, spatial frequency, and fractal dimension; The difference prediction model includes: The local feature extraction branch is used to process the local multidimensional statistical features through a multilayer perceptron to generate local descriptive factors; The global feature extraction branch is used to process the multi-channel feature map through the Transformer module to generate a global descriptor. The feature fusion and regression module is used to fuse the local descriptive factors and the global descriptive factors, and to regress and output the predicted inter-class differences. The difference prediction model was obtained through the following supervised training method: Obtain a training set of remote sensing images with pixel-level annotations and extract their multi-channel feature maps; For the multi-channel feature map, the target region and background region on the feature map are mapped according to the annotation of the corresponding original image. The feature mean of the two regions on the feature map is calculated respectively, and the norm between the two feature means is calculated as the actual inter-class difference ground truth label of the feature map. Using the multi-channel feature map as input and the actual inter-class difference ground truth labels as supervision targets, the mean squared error loss function is used to train the model end-to-end.
2. The feature enhancement method for offshore target detection according to claim 1, characterized in that, The adaptive sampling step S4 includes the following steps: S41: Obtain the initial candidate sampling point positions of the deformable convolution kernel; S42: Calculate the difference between each candidate sampling point and the center point; S43: Compare the inter-point difference value with the inter-class difference prediction value, and filter out candidate sampling points whose inter-point difference value is less than the inter-class difference prediction value to form a homogeneous sampling point set; S44: Determine whether the number of sampling points in the homogeneous sampling point set meets the preset convolution kernel size requirement; if not, sort the remaining candidate sampling points from the remaining candidate sampling points whose inter-point difference value is greater than or equal to the inter-class difference prediction value according to the size of their inter-point difference value, and add the candidate sampling points with the smallest inter-point difference value to the homogeneous sampling point set in turn until the quantity requirement is met, forming the final sampling point set. S45: Perform deformable convolution operation based on the final set of sampling points.
3. The feature enhancement method for offshore target detection according to claim 1, characterized in that, Step S5 includes the following steps: S51: Based on the offset of the final set of sampling points relative to the standard grid, the feature map is divided into edge regions and non-edge regions; S52: Convolve the edge region using the final set of sampling points to extract edge features; S53: Convolution is performed on non-edge regions using standard grid sampling points to extract internal texture features; S54: The edge features and the internal texture features are fused to generate an enhanced feature map, which is then used by the target detection network to obtain the detection result.
4. A feature enhancement system for maritime target detection, characterized in that The feature enhancement system implementing the method of claim 1 includes: The image processing module is used to acquire and preprocess remote sensing images containing maritime targets; The feature extraction module, connected to the image processing module, is used to extract multi-channel feature maps from the preprocessed image; The difference prediction module is connected to the feature extraction module and has a pre-trained difference prediction model embedded in it. It is used to receive the multi-channel feature map and output the inter-class difference prediction value. An adaptive sampling module is connected to the feature extraction module and the difference prediction module respectively, and is used to filter sampling points according to the inter-class difference prediction value to form a final sampling point set. The target detection module, connected to the adaptive sampling module, is used to extract and fuse features based on the final set of sampling points to perform maritime target detection.
5. A computer device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-4. The processor executes the computer program to implement the steps of the method according to any one of claims 1-3.
6. A computer-readable storage medium having stored thereon a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-3.
7. A computer program product comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-3.