An enhanced target detection method for complex beach scenes

Abstract

The application discloses an enhanced target detection method for a complex beach scene, comprising: acquiring an original beach scene image, and performing defogging and contrast enhancement processing on the original beach scene image to output a defogging image; performing semantic segmentation processing on the defogging image to obtain a semantic segmentation result of a beach target; performing channel dimension splicing on the defogging image and the semantic segmentation result to form a multi-channel input tensor; inputting the input tensor into a target detection network to complete target detection and complete enhanced detection of the target under the complex beach scene. The application can effectively improve the problem of blurred and low-contrast beach images, overcome the defect that the traditional segmentation method is sensitive to the direction of the target, improve the accuracy, robustness and environmental adaptability of target detection under the complex beach scene, and solve the problem of performance decline of the existing method under the conditions of strong noise interference, large scale change and variable target direction.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing technology, and specifically to an enhanced target detection method for complex beach scenes. Background Technology

[0002] Target detection in coastal scenes is a key technical task in applications such as marine environmental monitoring, nearshore security, and disaster assessment. However, due to the complexity of the natural environment, acquired coastal images often suffer from severe degradation, posing a significant challenge to existing general-purpose target detection methods. First, coastal areas are frequently affected by sea fog, suspended water vapor particles, and variable lighting, resulting in low contrast, blurred details, and color distortion in the captured images. While traditional or general-purpose image enhancement algorithms can improve the overall visual effect, they often lack targeted optimization for subsequent high-level visual tasks. During enhancement, they tend to weaken target edges or introduce irrelevant noise, leading to a disconnect between the preprocessing stage and the core recognition task.

[0003] Secondly, the orientation of targets on the shore (such as boats, floating objects, and facilities) is completely arbitrary. Current mainstream semantic segmentation and object detection models based on deep convolutional neural networks are highly dependent on the common orientations of targets in the training data and lack robustness to rotational changes. When a target rotates at a large angle, the features extracted by the model will change significantly, resulting in missed detections or false positives for the same category.

[0004] Furthermore, the beach scene encompasses both large-scale near-shore targets and extremely small-scale targets in the open sea, covering an extremely wide scale range. Simultaneously, the background contains highly interfering elements such as waves, cloud shadows, and sand textures. This makes it difficult for a single detection model to maintain stable performance under such extreme scale variations and complex noise backgrounds. Deep learning-based methods are sensitive to small targets and partially occluded targets, while methods based on traditional handcrafted features have limited discriminative power against complex texture backgrounds; both have their own limitations.

[0005] Current research and technologies largely focus on solving one of the aforementioned problems independently, such as designing dedicated dehazing algorithms, improving network structures to enhance scale adaptability, or introducing attention mechanisms. However, these solutions often fail to address the fundamental challenges of poor image quality, arbitrary target orientation, and multi-scale noise environments at the system level. Therefore, there is an urgent need in this field for a technical solution that can integrate and coordinate optimization from image preprocessing to target recognition to achieve robust and accurate target detection in complex beach scenes. Summary of the Invention

[0006] The purpose of this invention is to provide an enhanced target detection method for complex beach scenes, addressing the significant performance degradation of existing technologies in complex beach environments caused by image degradation, arbitrary target orientation, and varying scales. This method, through an integrated processing flow, collaboratively optimizes image preprocessing, feature representation, and target recognition, aiming to improve robustness against haze interference, stability in identifying targets in arbitrary orientations, and accuracy in multi-scale target detection, thereby achieving high-precision and high-reliability target detection in various beach scenes.

[0007] The technical solution to achieve the purpose of this invention is as follows:

[0008] An enhanced target detection method for complex beach scenes includes the following steps:

[0009] Step 1: Obtain the original beach scene image, and perform dehazing and contrast enhancement processing on the original beach scene image to output the dehazed image;

[0010] Step 2: Perform semantic segmentation processing on the dehazed image to obtain the semantic segmentation result of the beach target;

[0011] Step 3: Concatenate the dehazed image with the semantic segmentation result along the channel dimension to form a multi-channel input tensor;

[0012] Step 4: Input the input tensor into the target detection network to complete the target detection.

[0013] Further: Step 1 includes the following steps:

[0014] Step 1.1: Obtain the RGB three-channel image and texture enhancement feature map of the original beach scene image, and concatenate the RGB three-channel image and texture enhancement feature map to obtain a four-channel input image;

[0015] Step 1.2: Perform convolution operation on the four-channel input image to obtain the initial feature map;

[0016] Step 1.3: Encode the initial feature map to obtain coded features. ;

[0017] Step 1.4: Based on the encoded features, obtain the global features, local features, and semantically saliency guidance map, respectively. The local features are then calibrated to obtain recalibrated local features;

[0018] Step 1.5: The global features and the recalibrated local features are fused and then decoded to obtain a decoded feature map;

[0019] Step 1.6: Concatenate the decoded feature map with the semantic saliency guide map, and perform multi-level downsampling and upsampling operations on the concatenation result to obtain multi-scale fusion features;

[0020] Step 1.7: Map and reconstruct the multi-scale fusion features to obtain a dehazed image.

[0021] Furthermore, step 1.4 includes the following steps:

[0022] Step 1.4.1: Based on the encoded features, extract global features and local features respectively;

[0023] Step 1.4.2: Based on the encoded features, a semantically salient guide map is obtained, and its calculation formula is as follows:

[0024] ,

[0025] in, This represents a 1×1 convolution operation; This represents the Sigmoid activation function;

[0026] Step 1.4.3: Perform saliency perception on the local features, and introduce a weight mask generated based on foreground saliency detection for feature modulation. Densely connect and fuse the output features of each layer with the output features of the preceding layers to obtain saliency-perceived features; where the k-th layer output features... The calculation formula is as follows:

[0027] ,

[0028] in, These represent the output features from layer 0 to layer (k-1), Let w represent the nonlinear transformation function corresponding to the k-th layer, w represent the weight mask, ⊙ represent element-wise multiplication, and [.] represent feature concatenation.

[0029] Step 1.4.4: Perform global average pooling on the saliency-aware features to obtain channel statistical vectors, then input them into the fully connected layer to generate channel weights. The channel weights Acting on the corresponding feature channel, outputting recalibrated local features. The calculation formula is as follows:

[0030] ,

[0031] in, Indicates local features.

[0032] Furthermore, step 2 includes the following steps:

[0033] Step 2.1: Divide the dehazed image into support images. and query images Two parts, and obtain supporting images. Support mask ;

[0034] Step 2.2, for , and Perform rotation transformations separately to obtain the set of images that support rotation. Query image rotation set and support for mask rotation sets The rotation transformation angle , This indicates that when the rotation transformation angle is 0° , and The result of the rotation transformation, , and These represent rotation transformations of 90° respectively. , and The result of the rotation transformation, , and These represent the rotation transformation angles of 180° respectively. , and The result of the rotation transformation, , and These represent the rotation transformation angles of 270° respectively. , and The result of the rotation transformation;

[0035] Step 2.3: Input the support image rotation set and the query image rotation set into the pre-trained feature extractor to obtain the support feature set. With query feature set ;in They are respectively Feature extraction results They are respectively Corresponding query feature map;

[0036] Step 2.4: Based on the support feature set and the support mask rotation set, generate prototype vectors corresponding to each rotation transformation angle through mask weighted pooling operation:

[0037] ,

[0038] in represents the prototype vector with a rotation transformation angle of r, and ⊙ represents element-wise multiplication;

[0039] Step 2.5: Set the query features... Each query feature map in the dataset is flattened in the spatial dimension to obtain a set of query feature vectors under the corresponding rotation transformation angle. Where M is the total number of spatial locations, The query feature map representing the rotation transformation angle r is in the th position. The query feature vectors are located at each spatial position; the cosine similarity between each query feature vector and the prototype vector at the corresponding rotation transformation angle is calculated to obtain the similarity map for each rotation transformation angle. , its first Value at each spatial location for:

[0040] ,

[0041] in, · Describes the L2 norm of a vector;

[0042] Step 2.6: Set up the similarity map The components are concatenated, and a Softmax operation is performed on the concatenated result along the channel dimension to generate a rotation weight matrix. ;

[0043] Step 2.7: Based on the rotation weight matrix For the set of prototype vectors Weighted aggregation is performed to obtain rotationally robust fusion features. The calculation formula is as follows:

[0044] ,

[0045] in, This represents a vector concatenation operation;

[0046] Step 2.8: Incorporate the rotationally robust fusion feature With the query feature set Each query feature map in the dataset is concatenated, and a convolution operation is performed on the concatenated features to obtain enhanced query features for each rotation transformation angle. The calculation formula is as follows:

[0047] ,

[0048] in, Indicates the convolution operation;

[0049] Step 2.9: Enhance the query features for each rotation transformation angle. Input the void space pyramid pooling module and the prediction layer respectively to obtain semantic segmentation results for each rotation transformation angle. ;

[0050] Step 2.10: Semantic segmentation results for each rotation transformation angle The data is then fused to output the semantic segmentation results of the beach target.

[0051] Furthermore, the target detection network in step 4 includes a backbone detection path and an auxiliary detection path, and its processing steps include:

[0052] Step 4.1: Input the input tensor into the main detection path and the auxiliary detection path respectively to obtain two target detection results;

[0053] Step 4.2: Fuse the two target detection results to obtain the enhanced target detection result, thus completing the target detection.

[0054] Furthermore, in step 4.1, the backbone detection path is an improved YOLOv5+ model, which is obtained by improving the YOLOv5 model in the following way:

[0055] (a) Expand the width of the backbone network of the YOLOv5 model and replace the original Focus module with a convolutional layer;

[0056] (b) Replace the pre-selected convolutional module for feature extraction in the YOLOv5 model with a hybrid dilated residual module;

[0057] (c) Change the target matching strategy in the YOLOv5 model based on a fixed set of prior boxes and a standard intersection-union ratio (IU / U) metric to a target matching strategy based on an extended set of prior boxes and an improved IU / U metric metric;

[0058] (d) Change the feature fusion method in the YOLOv5 model to feature fusion that combines the semantic segmentation results with a multi-scale attention mechanism;

[0059] (e) The detection method in the YOLOv5 model that uses a single detection head to perform bounding box regression and class discrimination simultaneously is replaced with a task-decoupled detection head that uses a two-stage mechanism to handle the bounding box regression task and the class discrimination task separately.

[0060] Furthermore, the output characteristics of the hybrid void residual module The calculation formula is:

[0061] ,

[0062] ,

[0063] in, This represents the output of dilated convolution. This represents the input of the hybrid void residual module. This represents the weighting coefficient corresponding to the void ratio d. Indicates the maximum void ratio. This represents a dilated convolution operation with a dilation rate of d.

[0064] Furthermore, in step (c):

[0065] The original three prior boxes in the fixed prior box set are expanded to five prior boxes to obtain the extended prior box set.

[0066] The improved formula for calculating the intersection-union ratio (IUGR) is as follows:

[0067] ,

[0068] in, This indicates an improved intersection-union ratio metric. Indicates the prior box. Represents the true bounding box. Represents the balance factor. This represents the intersection operation. This represents the union operation. This indicates the area to be measured.

[0069] Furthermore, the feature fusion using the multi-scale attention mechanism combined with the semantic segmentation result in step (d) is calculated using the following formula:

[0070] ,

[0071] ,

[0072] ,

[0073] in, This represents element-wise multiplication. This represents the feature fusion result of the multi-scale attention mechanism that combines the semantic segmentation result. This represents the input for the feature fusion of the multi-scale attention mechanism that combines the semantic segmentation results. Indicates channel attention. Indicates spatial attention. represents the Sigmoid activation function, MLP represents the multilayer perceptron, GAP represents global average pooling, and GMP represents global max pooling. This represents a 7×7 convolutional layer. Indicates average pooling. This indicates max pooling. Indicates will and Then, the parts are assembled.

[0074] Furthermore, the task decoupling detection head used in step (e) to process the bounding box regression task and the category discrimination task separately through a two-stage mechanism includes the following processing steps:

[0075] In the first stage, a branch network is used to perform bounding box regression, which outputs the predicted location and size of the bounding box. The regression loss is CIoU loss.

[0076] In the second stage, another branch network is used for classification, outputting the class confidence of each predicted bounding box, and its classification loss adopts an improved Focal Loss.

[0077] Finally, the regression loss and the classification loss are superimposed as the loss function of the task decoupling head for training optimization.

[0078] Furthermore, the auxiliary detection path introduces a machine learning classifier based on multi-scale handcrafted feature detection, and its processing steps include:

[0079] (1) Obtaining visual saliency guidance maps using visual saliency detection algorithms The calculation formula is as follows:

[0080]

[0081] in Represents frequency domain variables, Represents the residuals of the logarithmic amplitude spectrum. Represents the phase spectrum. The imaginary unit, Indicates the inverse Fourier transform;

[0082] (2) Based on the aforementioned visual salience guidance map Multiple candidate regions are extracted from the dehazed image through threshold segmentation and connected component analysis. The spatial extent of each candidate region in the original image can be represented by a rectangular bounding box, and the local image corresponding to each candidate region is denoted as... , where k represents the candidate region index;

[0083] (3) For the local image corresponding to each candidate region Construct the Gaussian pyramid;

[0084] (4) On each scale layer of the Gaussian pyramid, extract SIFT descriptors for the corresponding candidate regions, and encode the features of the local sub-regions inside the candidate regions by dividing the spatial grids at different scales (the local sub-regions are small regions after dividing each candidate region), and generate corresponding multi-scale local feature vectors.

[0085] (5) The multi-scale local feature vectors obtained by each candidate region at different scales are concatenated to form a high-dimensional feature matrix, and the high-dimensional feature matrix is ​​subjected to feature encoding and dimensionality reduction processing by sparse dictionary learning method to obtain discriminative features;

[0086] (6) Input the discriminative features into a support vector machine classifier to perform binary classification of the candidate regions as targets and non-targets, and output the corresponding classification decision function value. The classification decision function value is normalized and used as the confidence level of the candidate region belonging to the target category, thereby obtaining the detection result of the auxiliary detection path, wherein each detection result includes the predicted bounding box corresponding to the candidate region and its confidence level.

[0087] Furthermore, in step 4, the two target detection results are fused, and the processing procedure includes:

[0088] ① Perform unified coordinate mapping on the predicted bounding boxes in the two target detection results, and match them based on the degree of overlap between the predicted bounding boxes, wherein the degree of overlap is measured by the Intersection over Union (IoU), and its calculation formula is as follows:

[0089] ,

[0090] in and These represent the predicted bounding boxes from the two detection pathways, This represents the area of ​​the region. When the IoU value between two predicted bounding boxes is greater than a preset threshold, the two predicted bounding boxes are considered to have matched successfully.

[0091] ② For the matched predicted bounding boxes, they are weighted and fused according to their respective class confidence, and redundant detection boxes are removed by non-maximum suppression to obtain the final enhanced target detection result.

[0092] Compared with the prior art, the significant advantages of this invention are:

[0093] 1. In the image dehazing stage, an innovative structure is adopted that integrates texture channel input with dual-stream collaborative processing. A four-channel input is constructed using extracted texture enhancement channels, and a global dehazing stream and a local enhancement stream based on a saliency-aware module are deployed in parallel within the encoder-decoder framework. This design not only effectively eliminates large-scale haze but also specifically restores and enhances texture details and edge structures in key areas, providing input images with significantly improved quality and greater task adaptability for subsequent segmentation and detection tasks.

[0094] 2. In the semantic segmentation stage, a robust prototype matching mechanism for rotated targets is proposed. This mechanism involves systematically rotating the support and query images from multiple angles, extracting prototypes in each direction, calculating dense similarities, and finally fusing multi-view features through an adaptively generated rotation weight matrix. This method enables the model to overcome the challenge posed by the arbitrary orientation of beach targets, achieving accurate and consistent segmentation of target contours with different orientations, and providing stable spatial prior information for subsequent detection.

[0095] 3. In the deep learning detection pathway, the YOLOv5+ model underwent comprehensive deep optimization. The proposed YOLOv5+ model integrates five core improvements: expanding the network width and replacing the Focus module to enhance basic feature capabilities; introducing a hybrid dilated residual module to expand the receptive field and capture multi-scale context; and improving the prior box matching degree calculation. The improvements include: a multi-scale attention module designed to leverage the interaction of high- and low-level features to enhance key regions; and task decoupling to separate regression and classification tasks to reduce interference. These improvements work synergistically to significantly enhance the detection performance for multi-scale targets, especially small targets.

[0096] 4. An auxiliary detection pathway based on traditional handcrafted features was constructed as an important supplement. This pathway uses the spectral residual method for visual saliency detection to extract candidate regions, then uses multi-scale SIFT feature descriptors combined with sparse dictionary learning for dimensionality reduction, and finally performs classification using support vector machines. This pathway is specifically designed for noisy environments and extreme-scale targets. Its stability and interpretability effectively compensate for the potential shortcomings of deep learning models in complex scenes, forming a deep integration with traditional methods.

[0097] 5. An integrated processing framework of "image enhancement—semantic segmentation—dual-stream detection" was constructed. This framework is not simply a series of three independent modules, but rather a design that enables deep collaboration between the preceding and following stages. For example, the dehazing module focuses on preserving structural information useful for segmentation and detection; the precise contours provided by the semantic segmentation module can be used to guide the focusing of the detection region; and the dual-stream detection pathway is enhanced from both data-driven and feature engineering perspectives. This systematic optimization strategy from preprocessing to advanced understanding specifically and comprehensively addresses the intertwined core challenges of beach scene images, such as blurriness, variable target orientation, and large scale differences. Attached Figure Description

[0098] Figure 1 This is a flowchart of the enhanced target detection method for complex beach scenes according to the present invention. Detailed Implementation

[0099] To better understand the present invention, the following description, in conjunction with the accompanying drawings, will provide further details.

[0100] Combination Figure 1 This invention provides an enhanced target detection method for complex beach scenes, comprising the following steps:

[0101] Step 1: Obtain the original beach scene image and input the original beach scene image into the image enhancement module to perform dehazing and contrast enhancement processing on the original beach image, and output the dehazed image;

[0102] This step addresses the issues of haze interference and low contrast in the original beach image by using specialized image enhancement processing to effectively improve image clarity and detail, providing a higher-quality input image for subsequent semantic segmentation and object detection modules. The specific steps are as follows:

[0103] Step 1.1: Input the beach image to be processed, and construct the input based on the input beach image to obtain a four-channel input image;

[0104] This step is to enable the method to utilize both color information and enhanced structural information in subsequent processing. The specific steps are as follows:

[0105] Step 1.1.1: Input the beach image to be processed, and extract the RGB three-channel data from the image;

[0106] Step 1.1.2: Perform LOG operator operations on the beach image to obtain a texture enhancement feature map;

[0107] Step 1.1.3: Perform channel concatenation between the RGB three-channel image and the texture enhancement feature map to obtain a four-channel input image.

[0108] Step 1.2: Input the four-channel input image into the initial convolution module, process the input image using 7×7 convolution operation, and output the initial feature map;

[0109] This step enables preliminary feature transformation and expands the receptive field, outputting an initial feature map.

[0110] Step 1.3: Input the initial feature map into the encoder network, and extract features through multi-layer convolution and downsampling operations to obtain the encoded features. ;

[0111] This step extracts deep semantic features through layer-by-layer convolution and nonlinear activation, and gradually expands the receptive field and compresses the spatial size through downsampling operations, ultimately obtaining encoded features containing high-level semantic information.

[0112] Step 1.4: Input the encoded features into the global processing stream and the local enhancement processing stream respectively to obtain the global features and the recalibrated local features, thus obtaining the semantic saliency guidance map. ;

[0113] The purpose of this step is to achieve the coordinated extraction and highlighting of overall structural information and key salient region information in the beach image, providing guidance for subsequent feature restoration and reconstruction. The specific steps are as follows:

[0114] Step 1.4.1: Input the encoded features into the global processing stream and the local enhancement processing stream respectively to obtain global features and local features, thus obtaining a semantically salient guidance map. The calculation formula is as follows:

[0115] ,

[0116] in: For encoding features; This represents a 1×1 convolution operation; This represents the Sigmoid activation function; A semantically significant guiding graph;

[0117] Step 1.4.2: Input the local features into the saliency perception module for layer-by-layer processing, and introduce a weight mask generated based on foreground saliency detection (FSD) during the feature transfer process. Feature modulation is performed, and the output features of each layer are densely connected and fused with the output features of the preceding layers to obtain saliency-aware features. The formula for calculating the layer's output is as follows:

[0118]

[0119] in, This represents the output characteristics of layer k. This represents the output features from layer 0 to layer (k-1). The nonlinear transformation function corresponding to the k-th layer is represented by , which consists of convolution, normalization, and activation operations; w represents the weight mask, and ⊙ represents element-wise multiplication, i.e., the weights and features are multiplied at corresponding positions. [] indicates feature splicing;

[0120] Step 1.4.3: Perform global average pooling on the saliency-aware features to obtain channel statistical vectors, input them into a fully connected layer to generate channel weights, apply the channel weights to the corresponding feature channels, and output recalibrated local features. The calculation formula is as follows:

[0121] ,

[0122] in, Indicates the local features to be recalibrated. ∠ represents local features, β represents channel weights, ⊙ represents element-wise multiplication, and ∠ represents the multiplication of weights with the feature map at corresponding positions.

[0123] Step 1.5: After fusing the global features and the recalibrated local features, input them into the decoder network, restore the feature resolution through stepwise upsampling and convolution operations, and output the decoded feature map;

[0124] The purpose of this step is to avoid oversmoothing or local distortion caused by relying on a single feature, and to provide a structurally consistent high-level representation for subsequent fine reconstruction.

[0125] Step 1.6: Combine the decoded feature map with the semantically saliency guide map. The features are then stitched together, and multi-level downsampling and upsampling operations are performed on the stitched features to obtain multi-scale fused features.

[0126] The purpose of this step is to use saliency maps to guide the decoding features at multiple scales, enabling the network to focus on key regions and background regions at different scales simultaneously, thereby reducing the risk of fog residue and loss of detail.

[0127] Step 1.7: Map and reconstruct the multi-scale fusion features to output a dehazed image;

[0128] Step 2: Input the dehazed image into the semantic segmentation module and process it using a rotation-robust semantic segmentation method to obtain the semantic segmentation result of the beach target;

[0129] The purpose of this step is to introduce a few-sample rotation robust semantic segmentation mechanism based on the dehazed image, making full use of limited annotation information to achieve stable and accurate segmentation of the target region even when the target is arbitrarily distributed in the direction of the beach, thus providing reliable spatial prior constraints for subsequent target detection. The specific steps are as follows:

[0130] Step 2.1: Obtain large-scale basic category sample data, pre-train the feature extraction network, and obtain a pre-trained feature extractor;

[0131] This step is solely for the purpose of pre-training the feature extraction network with a large number of basic category samples, enabling the network to have a general representation ability for common texture structures, edge contours and local morphological features in beach scenes, thereby improving the stability and generalization ability of feature extraction under conditions of few samples.

[0132] Step 2.2: Input the dehazed images output from Step 1 into the semantic segmentation module, selecting a portion of the dehazed images as support images. Obtain its support mask The remaining dehazed images are used as query images. The support image and the query image are rotated respectively, including rotations of 0°, 90°, 180° and 270°, to obtain a set of rotated support images. Query image rotation set Obtain the set of supported mask rotations The rotation transformation angle , This indicates that when the rotation transformation angle is 0° , and The result of the rotation transformation, , and These represent rotation transformations of 90° respectively. , and The result of the rotation transformation, , and These represent the rotation transformation angles of 180° respectively. , and The result of the rotation transformation, , and These represent the rotation transformation angles of 270° respectively. , and The result of the rotation transformation;

[0133] This step introduces multi-angle rotation transformation to explicitly simulate the appearance changes of beach targets in different directions, providing a foundation for subsequent rotation-robust feature modeling.

[0134] Step 2.3: Input the support image rotation set and the query image rotation set into the pre-trained feature extractor to obtain the support feature set. With query feature set ;in They are respectively Feature extraction results They are respectively Feature extraction results;

[0135] This step utilizes a pre-trained feature extractor to encode features of the support and query images at different rotation angles, mapping the original images to a high-dimensional semantic feature space, providing discriminative feature representations for prototype construction and similarity calculation.

[0136] Step 2.4: Based on the support feature set and the support mask rotation set, generate prototype vectors corresponding to each rotation angle through mask weighted pooling operation. The calculation formula is as follows:

[0137] ,

[0138] in This represents the prototype vector with a rotation transformation angle of r, and ⊙ represents element-wise multiplication. This step constructs a direction-related prototype vector that can represent the semantics of the beach target by aggregating only the feature information corresponding to the target region, thereby reducing the impact of background interference on the prototype representation.

[0139] Step 2.5: Set the query features... Each query feature map in the dataset is flattened in the spatial dimension to obtain a set of query feature vectors at the corresponding rotation angle r. Where M is the total number of spatial locations, , The query feature map with rotation angle r is in the th position. For each feature vector at a spatial location, calculate the prototype vector at the corresponding rotation angle for each query feature vector. The cosine similarity between them is used to obtain the similarity map for each rotation angle r. The value at each spatial location j for:

[0140] ,

[0141] in, · This represents the L2 norm of the vector. This step enables a fine-grained characterization of the semantic matching relationship between each spatial location in the query image and the target.

[0142] Step 2.6: Create similarity maps corresponding to each rotation angle. The feature maps are concatenated, and a Softmax operation is performed on the channel dimension of the concatenated feature maps to generate a rotation weight matrix. ;

[0143] Step 2.7: Based on the rotation weight matrix For each of the aforementioned prototype vectors Weighted aggregation is performed to obtain rotationally robust fusion features. The calculation formula is as follows:

[0144] ,

[0145] in, This represents the vector concatenation operation; this step eliminates feature differences caused by changes in target orientation by weighted fusion of multi-directional prototype features, constructs a target semantic representation that is insensitive to rotational changes, and improves the robustness of semantic segmentation in complex orientation distribution scenarios.

[0146] Step 2.8: Incorporate the rotationally robust fusion feature Compared with the original direction query features in the query feature set The features are concatenated, and a convolution operation is performed on the concatenated features to obtain the enhanced query features. The calculation formula is as follows:

[0147] ,

[0148] The purpose of this step is to enhance the semantics of the query features by injecting direction-invariant semantic prototype information into the query features, so that the features simultaneously contain local detail information and global target priors, thereby improving the ability to distinguish between segmentation boundaries and target regions.

[0149] Step 2.9: Enhance the query features from each rotational viewpoint. Input the Spatial Pyramid Pooling Module (ASPP) and the prediction layer respectively to obtain the semantic segmentation results for each rotation direction. , .

[0150] This step enhances the model's ability to perceive beach targets at different scales and generates semantic segmentation prediction results from each rotational viewpoint.

[0151] Step 2.10: Perform fusion processing on the semantic segmentation results corresponding to each rotation direction, and output the final semantic segmentation result of the beach target.

[0152] This step integrates the predictive advantages of different directions, suppresses possible missegmentation or omissions that may occur under a single perspective, and obtains stable and consistent semantic segmentation output for beach targets.

[0153] Step 3: Concatenate the dehazed image with the semantic segmentation result along the channel dimension to form a multi-channel input tensor, and input the input tensor into the main detection path and the auxiliary detection path respectively to obtain two target detection results;

[0154] The purpose of this step is to jointly model the dehazed image and the semantic segmentation result, and introduce a dual-path target detection structure. By utilizing the prior information of the target region provided by the semantic segmentation result, the generation of target candidate regions, feature extraction and localization discrimination process are guided and constrained, thereby achieving more robust target detection under the conditions of strong interference from complex beach backgrounds and large changes in target scale.

[0155] The specific steps are as follows:

[0156] Step 3.1: The dehazed image output from Step 1 and the semantic segmentation result output from Step 2 are concatenated along the channel dimension to form a multi-channel input tensor, which is then input to the target detection module with a dual-path structure. The semantic segmentation result serves as guiding information to constrain and enhance the target recognition and localization processing of the main detection path and the auxiliary detection path, respectively.

[0157] This step integrates semantic segmentation results with the dehazed image along the channel dimension, introducing explicit prior information about the target region. This enables the target detection module to distinguish between potential target regions and complex background regions during the feature learning stage, thereby suppressing the interference of irrelevant background features on the detection results and improving the reliability of target localization and classification.

[0158] Step 3.2: Process the input tensor through the backbone detection path to obtain the backbone detection result. The backbone detection path is based on the improved YOLOv5+ deep learning model.

[0159] The improvements made to the YOLOv5+ model include:

[0160] (3.2-a) The backbone network of the baseline YOLOv5 model is widened by 60%, and the original Focus module is replaced with a regular convolutional layer to enhance feature extraction capability and reduce structural complexity. The purpose of this improvement is to enhance the backbone network’s ability to represent complex beach textures and small-scale targets, while avoiding the loss of spatial information of small targets.

[0161] (3.2-b) The convolutional modules used for feature extraction in the baseline YOLOv5 model are improved by replacing them with hybrid dilated residual modules to expand the receptive field and fuse multi-scale contextual information. The output features... The calculation formula is:

[0162] ,

[0163] ,

[0164] in The void ratio, For the corresponding weighting coefficients, To achieve the maximum void ratio, ;

[0165] This improvement expands the network's receptive field without significantly increasing computational cost by introducing multi-void convolution and weighted fusion, enabling the model to simultaneously perceive local details and large-scale contextual information, thereby improving the stability of detecting beach targets at different scales.

[0166] (3.2-c) The target matching strategy in the baseline YOLOv5 model based on a fixed set of prior boxes and a standard intersection-union ratio (IU / R) metric is improved to a target matching strategy based on an expanded set of prior boxes and an improved IU / R metric to improve the accuracy of target matching in complex beach scenes. Specifically, the original three prior boxes are expanded to five, and the improved IU / R metric is used to calculate the prior box matching degree. The calculation formula is as follows:

[0167] ,

[0168] in For an improved crossover ratio metric, For prior boxes, For the true frame, As a balance factor, it is usually set to =0.5, This represents the intersection operation. This represents the union operation. This indicates the area to be measured;

[0169] This improvement enhances the matching accuracy between prior boxes and real targets in beach scenes with complex target scale distributions by expanding the number of prior boxes and introducing an improved cross-union ratio metric, thereby improving the recall capability for small and slender targets.

[0170] (3.2-d) The feature fusion method in the baseline YOLOv5 model, which does not incorporate semantic prior information, is improved to a multi-scale attention mechanism that combines the semantic segmentation results. This multi-scale attention module combines the semantic segmentation results and strengthens key region features through spatial and channel attention. The calculation formula is as follows:

[0171] ,

[0172] ,

[0173] ,

[0174] in This represents element-wise multiplication. This represents the enhanced attention feature map. Indicates input features, Indicates channel attention. Indicates spatial attention. (·) denotes the Sigmoid activation function, MLP represents a multilayer perceptron, typically consisting of two fully connected layers, GAP represents global average pooling, and GMP represents global max pooling. This represents a 7×7 convolutional layer used to fuse the concatenated feature maps. This indicates that average pooling will be performed on the input features. This indicates that max pooling will be performed on the input features. This indicates that the two pooling results will be concatenated.

[0175] This improvement uses a multi-scale attention module to guide the semantic segmentation results, jointly modeling channel attention and spatial attention, dynamically enhancing the feature response corresponding to the target region while suppressing background region features, thereby improving the detection model's attention to key regions and the efficiency of feature utilization.

[0176] (3.2-e) The detection method in the baseline YOLOv5 model, which uses a single detector head to perform bounding box regression and class discrimination simultaneously, is improved by using a two-stage mechanism to decouple the detection heads for the bounding box regression task and the class discrimination task respectively. The processing steps are as follows:

[0177] In the first stage, a branch network is used to regress bounding boxes on the feature map, outputting the predicted bounding box position and size. The regression loss adopts CIoU loss, and the loss function is designed as follows:

[0178] ,

[0179] in This represents the bounding box regression loss. and These are the center points of the predicted bounding box and the ground truth bounding box, respectively. For the improved intersection-union ratio (IUU) metric between predicted and ground truth boxes, ρ is the Euclidean distance, c is the diagonal length of the minimum closure region covering the two boxes, v is a term that measures aspect ratio similarity, and α is a weighting coefficient.

[0180] In the second stage, another branch network is used to classify the feature maps, outputting the class confidence score for each predicted bounding box. The classification loss adopts an improved Focal Loss, and the loss function is designed as follows:

[0181] ,

[0182] in, Represents the category classification loss. This represents the model's predicted probability of the true class. γ is the category balancing weight, and γ is the focusing parameter used to adjust the ratio of loss weights between difficult and easy samples;

[0183] Finally, the task decoupling head adopts the following loss function. Perform model training optimization:

[0184] ,

[0185] in For bounding box regression loss, The loss is categorized by class.

[0186] This improvement reduces feature conflicts between different tasks by decoupling bounding box regression from the class discrimination task, enabling the model to achieve both more accurate target localization and more stable class discrimination performance in complex beach backgrounds.

[0187] Step 3.3: The input tensor is processed in parallel through an auxiliary detection pathway to obtain auxiliary detection results. The auxiliary detection pathway is based on multi-scale handmade features and a machine learning classifier.

[0188] The purpose of this step is to provide a robust supplementary detection mechanism by introducing traditional detection methods based on handcrafted features, in situations where deep learning models may be affected by noise and extreme scale variations, thereby enhancing the overall reliability of the system in complex environments. The specific steps are as follows:

[0189] Step 3.3.1: Extract candidate regions by using a visual saliency detection algorithm based on spectral residuals to identify candidate regions from the dehazed image, wherein the saliency map... The calculation formula is:

[0190] ,

[0191] in Represents frequency domain variables, Represents the residuals of the logarithmic amplitude spectrum. Represents the phase spectrum. The imaginary unit, Indicates the inverse Fourier transform;

[0192] The purpose of this step is to quickly screen potential target regions through saliency detection, reduce the search space in the subsequent feature extraction and classification process, and improve the computational efficiency and anti-interference ability of the auxiliary detection pathway.

[0193] Step 3.3.2: Based on the aforementioned visual saliency guidance map Multiple candidate regions are extracted from the dehazed image through threshold segmentation and connected component analysis. The spatial extent of each candidate region in the original image can be represented by a rectangular bounding box, and the local image corresponding to each candidate region is denoted as... , where k represents the candidate region index;

[0194] Step 3.3.3: For each candidate region, examine the corresponding local image. Constructing a Gaussian pyramid for multi-scale sampling at different scales, its scale space function Defined as:

[0195] ,

[0196] Where G(x,y,σ)= , where is the Gaussian kernel function, and * denotes the convolution operation. For scale parameters;

[0197] Step 3.3.4: On each scale layer of the Gaussian pyramid, extract SIFT descriptors for the candidate regions, and encode the local sub-regions using 4×4 and 8×8 spatial grid partitioning methods respectively (the local sub-region is a small region after the partitioning of each candidate region), generating corresponding multi-scale local feature vectors;

[0198] The purpose of this step is to perform structured modeling of candidate regions through multi-scale SIFT feature description, enabling the auxiliary detection pathway to maintain strong invariance to the local texture and geometric structure of the target, thereby further improving detection stability.

[0199] Step 3.3.5: Concatenate the multi-scale local feature vectors obtained from each candidate region at different scales to form a high-dimensional feature matrix, and use the sparse dictionary learning method to perform feature encoding and dimensionality reduction on the high-dimensional feature matrix to obtain a compact discriminative feature representation;

[0200] This step can reduce feature dimensionality while retaining discriminative information, thereby improving the efficiency of classifier training and inference.

[0201] Step 3.3.6: Input the dimensionality-reduced discriminant features into a support vector machine classifier to perform binary classification of candidate regions as targets or non-targets, and output the corresponding classification decision function value. The classification decision function value is normalized and used as the confidence score of the candidate region belonging to the target category, thereby obtaining the detection result of the auxiliary detection path, wherein each detection result includes the predicted bounding box corresponding to the candidate region and its confidence score.

[0202] Step 4: Fuse the detection results from the dual-path target detection module and output the final target detection result to complete the enhanced target detection in complex beach scenes.

[0203] The purpose of this step is to fuse the detection results from the main detection pathway and the auxiliary detection pathway, combining the high expressive power of deep learning models with the stability advantages of traditional methods, effectively reducing false negatives and false positives, and outputting highly reliable target detection results suitable for complex beach scenes. The specific steps are as follows:

[0204] Step 4.1: Perform unified coordinate mapping on the predicted bounding boxes in the two detection results, and match the candidate detection results based on the degree of overlap between the predicted bounding boxes; wherein the degree of overlap is measured by the Intersection over Union (IoU), and its calculation formula is:

[0205] ,

[0206] in and These represent the predicted bounding boxes from the two detection pathways, This represents the area of ​​the region. When the IoU value between two predicted bounding boxes is greater than a preset threshold, the two predicted bounding boxes are considered to have matched successfully.

[0207] The purpose of this step is to achieve alignment and association of candidate targets from different detection paths in the same spatial reference system by uniformly mapping the predicted bounding boxes output by the main detection path and the auxiliary detection path, and matching them according to the degree of overlap between the bounding boxes, thus providing a consistent spatial basis for the subsequent fusion of detection results.

[0208] Step 4.2: For the matched predicted bounding boxes, perform weighted fusion based on their respective class confidence scores, and remove redundant detection boxes through non-maximum suppression to obtain the final enhanced target detection results.

[0209] In summary, this invention addresses the challenges of complex beach scenes, such as strong fog and haze interference, large target scale variations, irregular directional distribution, and severe background clutter. It constructs an enhanced target detection method that integrates image enhancement, semantic understanding, and target detection. This method first preprocesses the original beach image using a fusion-perception-guided image dehazing technique. This improves overall clarity while effectively preserving key structural and textural information, overcoming image blurring and low contrast, and providing high-quality input for subsequent processing. Then, a few-shot-based rotation-robust semantic segmentation technique is introduced. By performing rotation-adaptive matching and prediction fusion on targets in different directions, stable segmentation of beach targets in different directions is achieved under limited annotation conditions. This obtains semantic prior information with spatial constraints, achieving accurate and consistent segmentation of land feature outlines, effectively overcoming the sensitivity of traditional methods to target orientation. Based on this, the dehazed image and semantic segmentation results are jointly input into a dual-path target detection module. The main path is based on an improved YOLOv5+ model to enhance multi-scale feature representation and small target detection capabilities. The auxiliary path is based on a traditional target detection model with multi-scale SIFT features, utilizing handcrafted features to enhance the detection reliability in noisy environments and for targets at extreme scales. Through semantic prior guidance, the deep learning detection path works in conjunction with the traditional handcrafted feature detection path to achieve robust detection of multi-scale and multi-morphological targets, completing enhanced target detection in typical beach scenes. This method achieves coordinated optimization from low-level image quality improvement to high-level semantic understanding and then to target-level detection in the overall framework, effectively improving the accuracy, robustness and adaptability of target detection in complex beach scenes.

[0210] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. An enhanced target detection method for complex beach scenes, characterized in that, Includes the following steps: Step 1: Obtain the original beach scene image, and perform dehazing and contrast enhancement processing on the original beach scene image to output the dehazed image; Step 2: Perform semantic segmentation processing on the dehazed image to obtain the semantic segmentation result of the beach target; Step 3: Concatenate the dehazed image with the semantic segmentation result along the channel dimension to form a multi-channel input tensor; Step 4: Input the input tensor into the target detection network to complete target detection; Step 2 includes the following steps: Step 2.1: Divide the dehazed image into support images. and query images Two parts, and obtain supporting images. Support mask ; Step 2.2, for , and Perform rotation transformations separately to obtain the set of images that support rotation. Query image rotation set and support for mask rotation sets The rotation transformation angle , This indicates that when the rotation transformation angle is 0° , and The result of the rotation transformation, , and These represent rotation transformations of 90° respectively. , and The result of the rotation transformation, , and These represent the rotation transformation angles of 180° respectively. , and The result of the rotation transformation, , and These represent the rotation transformation angles of 270° respectively. , and The result of the rotation transformation; Step 2.3: Input the supporting image rotation set and the query image rotation set into the pre-trained feature extractor to obtain the supporting feature set. With query feature set ;in They are respectively Feature extraction results They are respectively Corresponding query feature map; Step 2.4: Based on the support feature set and the support mask rotation set, generate prototype vectors corresponding to each rotation transformation angle through mask weighted pooling operation: ， in represents the prototype vector with a rotation transformation angle of r, and ⊙ represents element-wise multiplication; Step 2.5: Set the query features... Each query feature map in the dataset is flattened in the spatial dimension to obtain a set of query feature vectors under the corresponding rotation transformation angle. Where M is the total number of spatial locations, The query feature map representing the rotation transformation angle r is in the th position. The query feature vectors are located at each spatial position; the cosine similarity between each query feature vector and the prototype vector at the corresponding rotation transformation angle is calculated to obtain the similarity map for each rotation transformation angle. , its first Value at each spatial location for: ， in, · Describes the L2 norm of a vector; Step 2.6: Set up the similarity map The components are concatenated, and a Softmax operation is performed on the concatenated result along the channel dimension to generate a rotation weight matrix. ; Step 2.7: Based on the rotation weight matrix For the set of prototype vectors Weighted aggregation is performed to obtain rotationally robust fusion features. The calculation formula is as follows: ， in, This represents a vector concatenation operation; Step 2.8: Incorporate the rotationally robust fusion feature With the query feature set Each query feature map in the dataset is concatenated, and a convolution operation is performed on the concatenated features to obtain enhanced query features for each rotation transformation angle. The calculation formula is as follows: ， in, Indicates the convolution operation; Step 2.9: Enhance the query features for each rotation transformation angle. Input the void space pyramid pooling module and the prediction layer respectively to obtain semantic segmentation results for each rotation transformation angle. ; Step 2.10: Semantic segmentation results for each rotation transformation angle Perform fusion processing to output the semantic segmentation results of the beach target; The target detection network in step 4 includes a backbone detection path and an auxiliary detection path, and its processing steps include: Step 4.1: Input the input tensor into the main detection path and the auxiliary detection path respectively to obtain two target detection results; Step 4.2: Fuse the two target detection results to obtain the enhanced target detection result, thus completing the target detection.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Obtain the RGB three-channel image and texture enhancement feature map of the original beach scene image, and concatenate the RGB three-channel image and texture enhancement feature map to obtain a four-channel input image; Step 1.2: Perform convolution operation on the four-channel input image to obtain the initial feature map; Step 1.3: Encode the initial feature map to obtain coded features. ; Step 1.4: Based on the encoded features, obtain the global features, local features, and semantically saliency guidance map, respectively. The local features are then calibrated to obtain recalibrated local features; Step 1.5: The global features and the recalibrated local features are fused and then decoded to obtain a decoded feature map; Step 1.6: Concatenate the decoded feature map with the semantic saliency guide map, and perform multi-level downsampling and upsampling operations on the concatenation result to obtain multi-scale fusion features; Step 1.7: Map and reconstruct the multi-scale fusion features to obtain a dehazed image.

3. The method according to claim 2, characterized in that, Step 1.4 includes the following steps: Step 1.4.1: Based on the encoded features, extract global features and local features respectively; Step 1.4.2: Based on the encoded features, a semantically salient guide map is obtained, and its calculation formula is as follows: ， in, This represents a 1×1 convolution operation; This represents the Sigmoid activation function; Step 1.4.3: Perform saliency perception on the local features, and introduce a weight mask generated based on foreground saliency detection for feature modulation. Densely connect and fuse the output features of each layer with the output features of the preceding layers to obtain saliency-perceived features; where the k-th layer output features... The calculation formula is as follows: ， in, These represent the output features from layer 0 to layer (k-1), Let w represent the nonlinear transformation function corresponding to the k-th layer, w represent the weight mask, and ⊙ represent element-wise multiplication. [] indicates feature splicing; Step 1.4.4: Perform global average pooling on the saliency-aware features to obtain channel statistical vectors, then input them into the fully connected layer to generate channel weights. The channel weights Acting on the corresponding feature channel, outputting recalibrated local features. The calculation formula is as follows: ， in, Indicates local features.

4. The method according to claim 1, characterized in that, In step 4.1, the backbone detection path is an improved YOLOv5+ model, which is obtained by improving the YOLOv5 model in the following way: (a) Expand the width of the backbone network of the YOLOv5 model and replace the original Focus module with a convolutional layer; (b) Replace the pre-selected convolutional module for feature extraction in the YOLOv5 model with a hybrid dilated residual module; (c) Change the target matching strategy in the YOLOv5 model based on a fixed set of prior boxes and a standard intersection-union ratio (IU / U) metric to a target matching strategy based on an extended set of prior boxes and an improved IU / U metric. (d) Change the feature fusion method in the YOLOv5 model to feature fusion that combines the semantic segmentation results with a multi-scale attention mechanism; (e) The detection method in the YOLOv5 model that uses a single detection head to perform bounding box regression and class discrimination simultaneously is replaced with a task-decoupled detection head that uses a two-stage mechanism to handle the bounding box regression task and the class discrimination task separately.

5. The method according to claim 4, characterized in that, In step (c): The original three prior boxes in the fixed prior box set are expanded to five prior boxes to obtain the extended prior box set. The improved formula for calculating the intersection-union ratio (IURR) is as follows: ， in, This indicates an improved intersection-union ratio metric. Indicates the prior box. Represents the true bounding box. Represents the balance factor. This represents the intersection operation. This represents the union operation. This indicates the area to be measured; The feature fusion using the multi-scale attention mechanism combined with the semantic segmentation result in step (d) is calculated using the following formula: ， ， ， in, This represents element-wise multiplication. This represents the feature fusion result of the multi-scale attention mechanism that combines the semantic segmentation result. This represents the input for the feature fusion of the multi-scale attention mechanism that combines the semantic segmentation results. Indicates channel attention. Indicates spatial attention. represents the Sigmoid activation function, MLP represents the multilayer perceptron, GAP represents global average pooling, and GMP represents global max pooling. This represents a 7×7 convolutional layer. Indicates average pooling. This indicates max pooling. Indicates will and Then, the parts are assembled.

6. The method according to claim 4, characterized in that, The two-stage mechanism used in step (e) to process the task decoupling detection head for bounding box regression and category discrimination tasks separately includes the following steps: In the first stage, a branch network is used to perform bounding box regression, which outputs the predicted location and size of the bounding box. The regression loss is CIoU loss. In the second stage, another branch network is used for classification, outputting the class confidence of each predicted bounding box, and its classification loss adopts an improved Focal Loss. Finally, the regression loss and the classification loss are superimposed as the loss function of the task decoupling detection head for training optimization.

7. The method according to claim 1, characterized in that, The auxiliary detection pathway involves introducing a machine learning classifier based on multi-scale handcrafted feature detection, and its processing steps include: (1) Obtaining visual saliency guidance maps using visual saliency detection algorithms The calculation formula is as follows: ， Where f represents frequency. Represents the residuals of the logarithmic amplitude spectrum. Represents the phase spectrum. The imaginary unit, Indicates the inverse Fourier transform; (2) Based on the aforementioned visual saliency guidance map Multiple candidate regions are extracted from the dehazed image through threshold segmentation and connected component analysis, and the local image corresponding to each candidate region is denoted as... , where k represents the candidate region index; (3) For the local image corresponding to each candidate region Construct the Gaussian pyramid; (4) On each scale layer of the Gaussian pyramid, extract SIFT descriptors for the corresponding candidate regions, and encode the sub-regions within the candidate regions using different scale spatial grid partitioning methods to generate corresponding multi-scale local feature vectors. (5) The multi-scale local feature vectors obtained by each candidate region at different scales are concatenated to form a high-dimensional feature matrix, and the high-dimensional feature matrix is ​​subjected to feature encoding and dimensionality reduction processing by sparse dictionary learning method to obtain discriminative features; (6) Input the discriminative features into the support vector machine classifier to perform binary classification of the candidate region as target and non-target, and output the corresponding classification decision function value. The classification decision function value is normalized and used as the confidence of the category to which the candidate region belongs. Finally, the detection result of the auxiliary detection path is obtained. The detection result of the auxiliary detection path includes the predicted bounding box corresponding to the candidate region and the corresponding confidence.

8. The method according to claim 1, characterized in that, Step 4 involves fusing the two target detection results, and the processing steps include: ① Perform unified coordinate mapping on the predicted bounding boxes in the two target detection results, and match them based on the degree of overlap between the predicted bounding boxes. If the degree of overlap between the two predicted bounding boxes is greater than a preset threshold, the two predicted bounding boxes are considered to be matched successfully; otherwise, the match is unsuccessful. ② For the successfully matched predicted bounding boxes, they are weighted and fused according to their respective class confidence scores, and redundant detection boxes are removed by non-maximum suppression to obtain the final enhanced target detection results.

Images