Foggy day long-distance ship target identification method based on multi-dimensional image cascade processing

Abstract

The invention provides a foggy day long-distance ship target identification method based on multi-dimensional image cascade processing, and relates to the technical field of computer vision and intelligent shipping monitoring, and the method comprises the steps: collecting a plurality of groups of marine ship images as an original data set, carrying out the downsampling operation of the original data set, generating a fogless ship image, and carrying out the downsampling operation of the fogless ship image; performing fog simulation on the fog-free ship image based on an atmospheric scattering model to obtain an initial foggy image; inputting the initial foggy image into an end-to-end defogging network, so that the end-to-end defogging network outputs a defogged original ship image; inputting the original ship image into a super-resolution reconstruction network to reconstruct a standard ship image; and inputting the standard ship image into a ship target detection network, and detecting a ship target in the standard ship image, so that the ship target detection network outputs a ship target identification result. According to the invention, the accuracy and stability of ship target detection and identification can be improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent shipping monitoring technology, and in particular to a method for long-distance ship target recognition in foggy weather based on multi-dimensional image cascade processing. Background Technology

[0002] With the rapid development of intelligent shipping and maritime supervision technologies, port and waterway monitoring systems are playing an increasingly important role in ship traffic safety management. Traditional video surveillance systems rely on optical imaging equipment to acquire ship target information, and can perform target detection and identification tasks well in clear, well-lit environments. However, in low-visibility weather conditions such as fog or haze, water vapor and particles suspended in the air cause severe scattering and absorption of visible light, resulting in problems such as low contrast, blurred details, and uneven grayscale distribution in the monitoring images. This leads to blurred edges and loss of contour information of ship targets, posing a significant challenge to the identification of long-distance targets.

[0003] Chinese patent CN118429810A discloses a method for ship target detection in foggy conditions based on mathematical data fusion, including the following steps: acquiring images of ships at sea under foggy conditions to be processed; performing image enhancement on the images of ships at sea under foggy conditions to be processed, and dividing the enhanced images into training and testing sets; establishing a ship detection network MDDShipNet, which includes an image processing filter model, a CNN-based filter parameter learner, and a YOLOv5M target detector based on a PSA attention module and a bidirectional pyramid WBiFPN feature fusion module; and inputting the test set data into the trained ship detection network MDDShipNet to achieve the detection of ships at sea in foggy scenes. However, the above scheme, in the detection of ships at long distances in foggy conditions, relies solely on feature fusion through the attention module and feature pyramid, resulting in insufficient perception of high-frequency detail features such as edges and textures of small targets at long distances, thus limiting detection accuracy. Therefore, it is necessary to provide a method for recognizing ships at long distances in foggy conditions based on multi-dimensional image cascade processing to improve the accuracy and stability of ship target detection and recognition. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for long-distance ship target recognition in foggy weather based on multi-dimensional image cascade processing.

[0005] This invention provides a method for long-range ship target recognition in foggy weather based on multi-dimensional image cascade processing, the method comprising: Multiple sets of images of ships at sea are collected as the original dataset. The original dataset is downsampled to generate fog-free ship images. Fog simulation is performed on the fog-free ship images based on an atmospheric scattering model to obtain initial foggy images. An initial foggy image is input into an end-to-end defogging network, and multi-scale features corresponding to the initial foggy image are extracted through the instance normalization layer, the embedded residual layer and the depth-separable convolutional layer in the end-to-end defogging network, so that the end-to-end defogging network outputs the original ship image after defogging. The original ship image is input into a super-resolution reconstruction network, and the feature representation of the ship target region in the original ship image is enhanced according to the attention mechanism of the super-resolution reconstruction network to reconstruct a standard ship image. The standard ship image is input into the ship target detection network to detect ship targets in the standard ship image, so that the ship target detection network outputs ship target recognition results, wherein the ship target recognition results include the category information and location information of each ship target.

[0006] Based on the above technical solutions, preferably, the end-to-end dehazing network includes an instance normalization layer, an embedded residual layer, and a depthwise separable convolutional layer connected in sequence, wherein, The instance normalization layer is used to receive the initial foggy image and normalize the features in the initial foggy image. The embedding residual layer and the instance normalization layer are connected through multiple convolutional layers. The embedding residual layer includes a first embedding residual block and a second embedding residual block. The first embedding residual block is used to receive the initial foggy feature map after preliminary convolution processing and extract the low-level texture features and edge information of the ship target to obtain the first foggy feature map. The second embedding residual block is used to receive the first foggy feature map and extract the high-level semantic features and complex fog and haze pattern features of the ship target to obtain the second foggy feature map. The embedded residual layer and the depthwise separable convolutional layer are connected by a first connection layer. The depthwise separable convolutional layer includes a first depthwise separable convolution, a second connection layer, a third connection layer, and a standard convolutional layer. The first depthwise separable convolution is used to receive the second foggy feature map and extract the spatial context information in the second foggy feature map through a combination of depthwise convolution and pointwise convolution. The second depthwise separable convolution is used to perform multi-scale feature extraction and optimization on the features fused by the second connection layer. The standard convolutional layer is used to receive the fused features output by the third connection layer and map the multi-dimensional features to the defogging parameter space through traditional convolution operations to generate the original defogging ship image.

[0007] Based on the above technical solutions, preferably, the super-resolution reconstruction network includes a first convolutional layer, a channel attention mechanism module, a spatial attention mechanism module, a second convolutional layer, a first activation function layer, a third convolutional layer, and a fusion unit, wherein, The first convolutional layer is used to perform an initial convolutional transformation on the received original ship image to extract shallow features of the original ship image and transmit the shallow features to the channel attention mechanism module; The channel attention mechanism module is used to receive the shallow features output by the first convolutional layer, generate channel features and channel attention weight vectors corresponding to the channel features, and transmit the weighted channel features to the spatial attention mechanism module. The spatial attention mechanism module is used to receive the weighted channel features, generate a spatial attention weight map and spatial attention features based on the weighted channel features, and transmit the spatial attention features to the fusion unit; The second convolutional layer is used to receive the original ship image, perform convolution processing on the original ship image to extract mid-level structural features, and transmit the mid-level structural features to the first activation function layer and the fusion unit; The first activation function layer is used to perform a nonlinear transformation on the mid-level structural features output by the second convolutional layer, and then transmits the nonlinear features obtained after the nonlinear transformation to the third convolutional layer. The third convolutional layer is used to extract the nonlinear features through convolution to obtain residual features, and the residual features, the spatial attention features, and the original ship image are reconstructed in the fusion unit to obtain a standard ship image.

[0008] More preferably, the spatial attention mechanism module includes an average pooling layer, a max pooling layer, a fourth connection layer, a fourth convolutional layer, and a second activation function layer, wherein, The average pooling layer is used to receive the weighted channel features output by the channel attention mechanism module, and to perform a global average pooling operation on the weighted channel features in the channel dimension to obtain a spatial statistical feature map. The max pooling layer is used to receive the weighted channel features output by the channel attention mechanism module, and to perform a global max pooling operation on the weighted channel features in the channel dimension to obtain a spatial saliency feature map. The fourth connection layer is used to concatenate the spatial statistical feature map and the spatial saliency feature map in the channel dimension to generate a spatial fusion feature map containing two channels. The fourth convolutional layer is used to receive the spatial fusion feature map and perform spatial convolution processing on the spatial fusion feature map through convolution operations to generate spatial attention features; The second activation function layer is used to perform nonlinear activation on the spatial attention features and normalize the spatial attention features to generate a spatial attention weight map.

[0009] More preferably, the ship target detection network includes a YOLOv10 network and an LSDECD detection head, wherein, The LSDECD detection head includes multiple input convolutional layers, a first detail enhancement convolutional layer, a second detail enhancement convolutional layer, multiple regression and classification convolutional groups, and multiple scaling layers, wherein... The input convolutional layer is used to receive ship feature maps of different scales extracted by the YOLOv10 backbone network; The first detail enhancement convolutional layer is used to fuse feature maps of different scales to generate a first multi-scale fusion feature with preliminary enhancement, and then transmits the first multi-scale fusion feature to the second detail enhancement convolutional layer. The second detail-enhancing convolutional layer is used to receive the first multi-scale fusion feature, generate a sufficiently enhanced second multi-scale fusion feature, and distribute the second multi-scale fusion feature to multiple parallel regression-classification convolutional groups; The multiple regression and classification convolutional groups correspond one-to-one with different detection scales. Each regression and classification convolutional group includes a parallel regression convolutional branch and a classification convolutional branch. The regression convolutional branch is used to perform convolution processing on the second multi-scale fusion feature to predict the bounding box coordinate parameters of the ship target. The classification convolutional branch is used to perform convolution processing on the second multi-scale fusion feature to predict the class probability distribution of the ship target. The multiple scaling layers are connected to the output of each regression convolution branch, and each scaling layer is used to perform scale adaptive adjustment and normalization on the bounding box prediction results of the corresponding scale.

[0010] More preferably, the first detail-enhancing convolutional layer includes a fifth convolutional layer, a central difference convolutional layer, an angular difference convolutional layer, a vertical difference convolutional layer, and a horizontal difference convolutional layer, wherein, The fifth convolutional layer is used to receive feature maps of different scales and form a reference feature corresponding to the ship feature map; The central difference convolutional layer is used to calculate the central difference value between each pixel in the ship feature map and its surrounding neighboring pixels, so as to extract radial gradient features centered on the pixel. The angular difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its diagonal neighboring pixels in order to extract gradient features in the diagonal direction. The vertical difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its vertically adjacent pixels above and below it, so as to extract the gradient features in the vertical direction. The horizontal difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its horizontal left and right neighboring pixels to extract the gradient features in the horizontal direction.

[0011] More preferably, the end-to-end defogging network, the super-resolution reconstruction network, and the ship target detection network are cascaded in a serial manner.

[0012] A second aspect of this application provides a foggy long-range ship target recognition system based on multi-dimensional image cascade processing. The foggy long-range ship target recognition system includes a data acquisition module, a data processing module, and a target recognition module, wherein... The data acquisition module is used to acquire multiple sets of images of ships at sea as the original dataset, perform downsampling on the original dataset to generate fog-free ship images, and simulate fog on the fog-free ship images based on an atmospheric scattering model to obtain an initial foggy image. The data processing module is used to input the initial foggy image into an end-to-end defogging network, and extract multi-scale features corresponding to the initial foggy image through the instance normalization layer, embedded residual layer and depth separable convolutional layer in the end-to-end defogging network, so that the end-to-end defogging network outputs the defogging original ship image. The original ship image is then input into a super-resolution reconstruction network, and the feature representation of the ship target region in the original ship image is enhanced according to the attention mechanism of the super-resolution reconstruction network to reconstruct a standard ship image. The target recognition module is used to input the standard ship image into the ship target detection network, detect the ship targets in the standard ship image, and enable the ship target detection network to output ship target recognition results, wherein the ship target recognition results include the category information and location information of each ship target.

[0013] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory.

[0014] A fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a method for identifying long-range ship targets in foggy weather based on multidimensional image cascade processing.

[0015] The foggy long-distance ship target recognition method and system based on multi-dimensional image cascade processing provided by this invention has the following advantages over existing technologies: (1) By downsampling the original marine ship images and combining them with atmospheric scattering models to simulate fog, a dataset of foggy images with different fog concentrations and imaging conditions can be generated while maintaining the consistency of ship structure and semantic information. This improves the algorithm's adaptability to complex sea conditions in real foggy weather and alleviates the problem of difficult collection of real foggy samples. Furthermore, by using an end-to-end defogging network containing instance normalization layers, embedded residual layers, and depth separable convolutional layers, multi-scale feature extraction and fusion can be performed on the initial foggy images. This can effectively weaken the influence of fog scattering, enhance image contrast and clarity, suppress noise amplification, and better preserve ship edges and texture details. The original ship images after defogging are input into the super-resolution reconstruction network. The attention mechanism is used to focus on enhancing the ship target area, which helps to highlight the key structural information of the ship, suppress background sea surface and sky interference, achieve high-quality reconstruction of distant and small-sized ship targets, and improve the overall resolution and target recognizability of the image. At the same time, by performing target detection on standard ship images after defogging and super-resolution reconstruction, the detection network can work under clearer and higher resolution input conditions, thereby significantly improving the accuracy and stability of ship target detection and recognition.

[0016] (2) By sequentially introducing instance normalization layer, embedded residual layer and depth separable convolutional layer into the end-to-end dehazing network, hierarchical modeling and efficient dehazing processing of features of ship images in foggy weather are realized. The depth separable convolutional layer, combined with a multi-level connection structure, fully explores spatial context information and realizes multi-scale feature fusion under the premise of reducing computational complexity and parameter scale. Finally, the fused features are mapped to the dehazing parameter space through standard convolution, thereby generating a dehazed ship image with clear details, enhanced contrast and significant fog suppression effect. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the foggy long-distance ship target recognition method based on multi-dimensional image cascade processing provided by the present invention. Figure 2 This is a schematic diagram of the end-to-end defogging network provided by the present invention; Figure 3A schematic diagram of the structure of the super-resolution reconstruction network provided by this invention; Figure 4 A schematic diagram of the spatial attention mechanism module provided by the present invention; Figure 5 This is a schematic diagram of the structure of the LSDECD detection head provided by the present invention; Figure 6 A schematic diagram of the structure of the first detail-enhancing convolutional layer provided by the present invention. Figure 7 This is a schematic diagram of the structure of the long-range ship target identification system in foggy weather provided by the present invention; Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0019] Explanation of reference numerals in the attached diagram: 1. Fog-weather long-range ship target identification system; 11. Data acquisition module; 12. Data processing module; 13. Target identification module; 2. Electronic equipment; 21. Processor; 22. Communication bus; 23. User interface; 24. Network interface; 25. Memory. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention discloses a method for long-range ship target recognition in foggy weather based on multi-dimensional image cascade processing, with reference to... Figure 1 The steps of this method include S1 to S4.

[0022] Step S1: Collect multiple sets of marine vessel images as the original dataset, perform downsampling on the original dataset to generate fog-free vessel images, and simulate fog on the fog-free vessel images based on an atmospheric scattering model to obtain initial foggy images.

[0023] In this step, visible light video surveillance cameras deployed at ports, waterways, shore-based radar stations, or coastlines continuously capture images of ships sailing or anchored at sea, obtaining images of ships at different distance scales, with different ship types and background conditions. Shipborne camera equipment, drones, or image acquisition devices mounted on patrol boats are used to capture images of ship targets in distant sea areas, obtaining multi-view, multi-resolution ship image data. The acquired ship images cover weather conditions without significant fog interference, such as sunny and cloudy days, and include scenes with different light intensities, sea surface reflections, and background complexity to ensure the diversity and representativeness of the original dataset.

[0024] The high-resolution images of ships at sea are downsampled according to a preset scaling ratio. For example, bilinear interpolation or bicubic interpolation algorithms are used to reduce the spatial resolution of the original image to the target resolution to simulate the low-resolution appearance of ship targets in the image under long-distance imaging conditions. During the downsampling process, the overall structural information of the image and the ship outline are kept from significant distortion to ensure that the downsampled image can still realistically reflect the appearance features of long-distance ship targets. The downsampled image is used as a fog-free ship image for subsequent fog simulation processing and network training sample construction.

[0025] After obtaining fog-free images of the ship, fog simulation is performed on these images based on an atmospheric scattering model to generate initial foggy images. The observed foggy images are then analyzed. This can be represented as a direct attenuation term (and the superposition of the atmospheric light scattering term):

[0026] in, Indicates the observed image in pixels Brightness at that location Indicated in pixels The actual radiation intensity in a fog-free environment. Indicates the global atmospheric light value. Indicates the atmospheric scattering coefficient. Indicated in pixels Scene depth at that location Indicated in pixels Transmittance is the proportion of light energy that passes through the atmosphere at a given location. Transmittance is related to scene depth. It exhibits an exponential decay relationship.

[0027] In one example, the atmospheric scattering coefficient can be set randomly or in segments. The range of values ​​can be determined to simulate fog scenarios with different concentrations and distribution patterns; simultaneously, global atmospheric light can be analyzed. Random sampling or preset fixed values ​​are used to enhance the diversity of generated foggy images. By substituting the fog-free ship image into the aforementioned atmospheric scattering model, and combining the set atmospheric light parameters and transmittance distribution, fog synthesis processing is performed on each pixel to generate an initial foggy image consistent with the imaging characteristics of real foggy days.

[0028] Furthermore, in actual video surveillance, scene depth It is difficult to obtain directly. To avoid explicit ranging, this invention utilizes a deep learning model. Perform transmittance prediction:

[0029] in, This represents the pixel position predicted by the deep learning model. Estimated transmittance at [location] This represents a learnable mapping function that includes a convolutional encoder and an attention fusion layer. Indicates input image with fog After normalization, at pixel position The pixel value at that location. The network uses physical consistency loss as a constraint during the training phase:

[0030] in, This represents the physical consistency loss function for transmittance prediction. Describing the L1 norm, Indicates pixel position The initial transmittance estimate at that location, This is the edge preservation coefficient. Indicates predicted transmittance At pixel position Spatial gradient at that location Indicates initial transmittance At pixel position The spatial gradient at that point. This loss encourages the network to predict transmittance in accordance with the physical model while maintaining edge continuity.

[0031] Traditional methods often take the area with the highest brightness in the entire image as the atmospheric light value. However, this method is prone to errors on highly reflective sea surfaces. Therefore, this application proposes an adaptive estimation based on region weighting: in, This represents the candidate region selected by the brightness threshold. This represents the local smoothing weight of a pixel.

[0032] Calculate the preliminary restored image: in, Indicates pixel position Preliminary restored image pixel values ​​at the location, This represents the minimum transmittance threshold. To prevent excessive noise amplification. A channel correction factor is introduced to avoid color deviation. :

[0033] in, This indicates the pixel position of the final restored image after channel correction. Color channels pixel values ​​on Indicates the preliminary restored image In color channels pixel values ​​on This represents the average channel values ​​of the haze-free reference image. Indicates the preliminary restored image In color channels The average pixel value on the surface.

[0034] To further improve illumination consistency, a method based on [previous method] was applied to the restoration results. Retinex Multi-scale illumination balance algorithm, restored reflection components After gamma correction: Step S2: Input the initial foggy image into the end-to-end defogging network, and extract multi-scale features corresponding to the initial foggy image through the instance normalization layer, the embedded residual layer and the depth separable convolutional layer in the end-to-end defogging network, so that the end-to-end defogging network outputs the original ship image after defogging.

[0035] In this step, to address the issues of low contrast and blurred details in ship images under foggy conditions, an end-to-end defogging network (AOD-Net) is employed. To improve the defogging effect of ship images in foggy conditions and the network's adaptive ability to fog concentration, each channel of each initially foggy image is independently normalized. The calculation formula is as follows: in, x This represents the initial foggy image as input. This represents the mean of the feature map across each channel. This represents the standard deviation of the feature map in each channel. It is a very small constant (usually 1e) -5 This is used to prevent the denominator from being 0. and These are trainable parameters. Introducing instance normalization (InstanceNorm2d) can improve the network's robustness to changes in the brightness and contrast of the input image, enabling the network to better adapt to different foggy scenes.

[0036] In this step, the end-to-end dehazing network comprises, in sequence, an instance normalization layer, an embedded residual layer, and a depthwise separable convolutional layer, wherein, The instance normalization layer is used to receive the initial foggy image and normalize the features in the initial foggy image; The embedding residual layer and the instance normalization layer are connected by multiple convolutional layers. The embedding residual layer includes a first embedding residual block and a second embedding residual block. The first embedding residual block is used to receive the initial foggy feature map after preliminary convolution processing and extract the low-level texture features and edge information of the ship target to obtain the first foggy feature map. The second embedding residual block is used to receive the first foggy feature map and extract the high-level semantic features and complex fog and haze pattern features of the ship target to obtain the second foggy feature map. The embedded residual layer and the depthwise separable convolutional layer are connected by a first connecting layer. The depthwise separable convolutional layer includes a first depthwise separable convolution, a second connecting layer, a third connecting layer, and a standard convolutional layer. The first depthwise separable convolution is used to receive the second foggy feature map and extract the spatial context information in the second foggy feature map through a combination of depthwise convolution and pointwise convolution. The second depthwise separable convolution is used to perform multi-scale feature extraction and optimization on the features fused by the second connecting layer. The standard convolutional layer is used to receive the fused features output by the third connecting layer. Through traditional convolution operations, the multidimensional features are mapped to the defogging parameter space to generate the original ship image after defogging.

[0037] In this embodiment, by sequentially introducing an instance normalization layer, an embedded residual layer, and a depthwise separable convolutional layer into the end-to-end dehazing network, hierarchical modeling and efficient dehazing of features in foggy ship images are achieved. The instance normalization layer effectively eliminates distribution differences caused by different imaging conditions and fog concentrations, stabilizes network training, and improves adaptability to complex foggy environments. The embedded residual layer, through the synergistic effect of the first and second embedded residual blocks, progressively extracts low-level texture and edge information as well as high-level semantics and complex fog patterns of ship targets, enhancing effective feature representation while avoiding the gradient vanishing problem caused by deep networks. Furthermore, the depthwise separable convolutional layer, combined with a multi-level connection structure, fully mines spatial context information and achieves multi-scale feature fusion while reducing computational complexity and parameter scale. Finally, the fused features are mapped to the dehazing parameter space through standard convolution, thereby generating a dehazed ship image with clear details, enhanced contrast, and significant fog suppression effect, providing high-quality input for subsequent high-precision detection and recognition of distant ship targets.

[0038] Furthermore, such as Figure 2 As shown, Figure 2 This diagram illustrates the structure of an end-to-end dehazing network. By introducing instance normalization, the network can better adapt to input images with varying fog concentrations and color deviations, thereby improving its robustness and generalization ability. To this end, the end-to-end dehazing network (AOD-Net) adds an instance normalization layer at the input, performing independent normalization processing on each channel of each input image. Introducing the instance normalization layer improves the network's robustness to changes in input image brightness and contrast, enabling it to better adapt to different foggy scenarios. This enhances the network's adaptability to varying fog concentrations and lighting conditions.

[0039] Secondly, embedding residual blocks deepens the network structure, improves feature extraction capabilities, effectively alleviates the gradient problem during deep network training, and replaces some standard convolutions with depthwise separable convolutions, significantly reducing the number of model parameters and computational complexity while maintaining the receptive field; the input feature maps of convolutional layers are typically dimensional... To indicate, among which, The number of channels in the input feature map. and These represent the width and height of the feature map, respectively. During a standard convolution operation, in order to generate a feature map with dimensions of... The output feature map needs to be applied to a set of sizes. The convolution operation is performed using the convolution kernel, where, The number of channels in the output feature map. Let be the size of the convolution kernel. Therefore, the computational complexity of a standard convolution operation is usually expressed as the total number of multiplications and additions. .

[0040] Compared to standard convolution operations, the computation process of depthwise separable convolution can be divided into two main steps: first, performing a depthwise convolution operation, followed by a pointwise convolution operation. The depthwise convolution operation uses a size of... The convolutional kernels process the input feature map according to the channel dimension, where each kernel has a number of channels. The convolution kernels operate on the corresponding input feature map channels, so a single depthwise convolution operation can be represented by the following formula:

[0041] in, Indicates the input feature map at the th c The first channel, width direction w The position, the height direction h Feature values ​​at each position, Represents the convolution kernel. u This indicates the first convolution kernel in the width direction. u coordinates, v This indicates the first convolution kernel in the height direction. v coordinates, Indicates the input feature map, This is the output feature map after depthwise convolution. c Represents the first... c One channel, This indicates the output feature map in the width direction. w Each coordinate position This indicates the output feature map in the height direction. h Each coordinate position.

[0042] Specifically, both DSConv1 (5×5×3) and DSConv2 (7×7×3) consist of two parts: depthwise convolution and pointwise convolution. While maintaining the receptive field size, this significantly reduces the number of network parameters and computational cost, improving network efficiency and facilitating practical deployment. For activation functions, all convolutional layers use the LeakyReLU activation function instead of the original AOD-Net's ReLU. The LeakyReLU activation function effectively avoids the neuron "death" phenomenon that can occur with negative inputs, allowing negative gradients to be effectively propagated, enhancing the network's nonlinear fitting ability, and contributing to a more precise modeling of the haze degradation process. In the feature fusion part of the end-to-end dehazing network, the AOD-Net's Concat stitching layer structure is still used to effectively fuse multi-scale features extracted from different convolutional layers to obtain richer contextual information. Finally, the stitched feature map is used to estimate the parameter K through a Conv5 (1×1×1) convolutional layer, resulting in a clear dehazed image of the ship. The entire end-to-end dehazing network takes foggy ship images as input and clear, fog-free images as the target output, and is optimized through a hybrid loss function (a weighted sum of MSE loss and perception loss).

[0043] Step S3: Input the original ship image into the super-resolution reconstruction network, and enhance the feature representation of the ship target region in the original ship image according to the attention mechanism of the super-resolution reconstruction network, so as to reconstruct the standard ship image.

[0044] In this step, the dehazed original ship image output from step S2 is used as input to the super-resolution reconstruction network. The original ship image is a low-resolution, clear image, with a spatial resolution lower than the preset standard ship image resolution.

[0045] A first convolutional layer is set at the front end of the super-resolution reconstruction network to perform convolutional transformation on the input original ship image and extract shallow feature information of the image. The shallow features mainly include the basic outline, edge structure and local texture features of the ship target, and serve as the basic feature input for subsequent attention enhancement processing.

[0046] The channel attention mechanism module obtains global statistical information for each channel by performing global average pooling and global max pooling operations on shallow features, and generates channel attention weight vectors through a non-linear mapping network with shared weights. Based on the features enhanced by the channel attention mechanism, the spatial attention mechanism module generates spatial feature maps by performing average pooling and max pooling operations along the channel dimension, and generates spatial attention weight maps through convolution operations and activation functions.

[0047] After attention enhancement, the enhanced features are fused with the original shallow features using residual connections to form a fused feature that includes both the original information and the attention-enhanced information. Through the residual learning mechanism, the overall structural information of the image can be preserved while enhancing the target features, avoiding detail distortion caused by over-enhancement.

[0048] In this step, the super-resolution reconstruction network includes a first convolutional layer, a channel attention mechanism module, a spatial attention mechanism module, a second convolutional layer, a first activation function layer, a third convolutional layer, and a fusion unit. The first convolutional layer is used to perform an initial convolutional transformation on the received original ship image to extract shallow features of the original ship image and transmit the shallow features to the channel attention mechanism module. The channel attention mechanism module is used to receive the shallow features output by the first convolutional layer, generate channel features and channel attention weight vectors corresponding to the channel features, and transmit the weighted channel features to the spatial attention mechanism module. The spatial attention mechanism module is used to receive weighted channel features, generate a spatial attention weight map and spatial attention features based on the weighted channel features, and transmit the spatial attention features to the fusion unit; The second convolutional layer is used to receive the original ship image, perform convolution processing on the original ship image to extract the mid-level structural features, and transmit the mid-level structural features to the first activation function layer and the fusion unit. The first activation function layer is used to perform nonlinear transformation on the mid-level structural features output by the second convolutional layer, and then transmits the nonlinear features obtained after the nonlinear transformation to the third convolutional layer. The third convolutional layer is used to extract nonlinear features through convolution to obtain residual features, and the residual features, spatial attention features, and original ship images are reconstructed in the fusion unit to obtain standard ship images.

[0049] Furthermore, the spatial attention mechanism module includes an average pooling layer, a max pooling layer, a fourth connection layer, a fourth convolutional layer, and a second activation function layer, wherein, The average pooling layer is used to receive the weighted channel features output by the channel attention mechanism module, and to perform global average pooling on the weighted channel features in the channel dimension to obtain a spatial statistical feature map. The max pooling layer is used to receive the weighted channel features output by the channel attention mechanism module, and to perform global max pooling operation on the weighted channel features in the channel dimension to obtain a spatial saliency feature map. The fourth connection layer is used to concatenate the spatial statistical feature map and the spatial saliency feature map along the channel dimension to generate a spatial fusion feature map containing the two channels. The fourth convolutional layer is used to receive the spatial fusion feature map, and performs spatial convolution processing on the spatial fusion feature map through convolution operation to generate spatial attention features; The second activation function layer is used to perform non-linear activation on the spatial attention features and normalize the spatial attention features to generate a spatial attention weight map.

[0050] In one example, such as Figure 3 and Figure 4 As shown, in terms of super-resolution reconstruction, improvements are made to the A2F-SR network by incorporating a CBAM attention module, enhancing the network's ability to reconstruct ship target details in low-resolution images. The input feature map is first processed by a channel attention mechanism module. This module uses global average pooling and global max pooling in parallel to aggregate global information from each channel, and then determines the contribution of each channel through a learning process. The calculated channel weights are multiplied back onto the original feature map to amplify the feature responses of important channels. Next, the channel-weighted feature map is fed into a spatial attention mechanism module. This module performs maximum and average operations along the channel dimension at each spatial location of the feature map and concatenates these two results to generate a spatial attention map.

[0051] Step S4: Input the standard ship image into the ship target detection network to detect the ship targets in the standard ship image, so that the ship target detection network outputs the ship target recognition result, wherein the ship target recognition result includes the category information and location information of each ship target.

[0052] In this step, the standard ship image output from step S3 is used as input data and fed into the ship target detection network. The standard ship image is a high-quality ship image that has undergone dehazing and super-resolution reconstruction, and its spatial resolution and sharpness meet the input requirements of the ship target detection network.

[0053] In the specific implementation process, the standard ship images can be normalized in size, aligned in channels, and normalized in pixel value to ensure that standard ship images from different sources have a consistent data format, thereby improving the stability and robustness of the ship target detection network during the training and inference phases.

[0054] The ship target detection network is a deep learning-based target detection network, comprising a backbone feature extraction network, a feature fusion network, and a detection head module. The detection network preferably uses the YOLOv10 network as its basic structure, combined with a detection head designed for the characteristics of long-distance ship targets, to achieve real-time detection of ship targets in standard ship images. The input standard ship image is fed into the backbone feature extraction network, where feature maps of different scales are extracted through multi-layer convolutional operations. These feature maps include high-resolution feature layers suitable for detecting small-sized ship targets, and low-resolution feature layers suitable for detecting medium- and large-sized ship targets. Subsequently, the feature maps of different scales are input into the feature fusion network for fusion processing to integrate multi-scale spatial and semantic information, improving the detection capability for small-sized ship targets at long distances.

[0055] The fused multi-scale features are input into the detection head module for target prediction. The detection head module includes multiple detection branches, each corresponding to a feature map at a different scale. Within each detection branch, ship targets are detected through parallel classification and regression prediction branches.

[0056] The classification prediction branch is used to output the category information of each candidate target in the standard ship image. The category information includes the category labels of different types of ships and their corresponding confidence scores. The regression prediction branch is used to output the location information of each candidate target, which includes the bounding box coordinate parameters of the ship target in the standard ship image.

[0057] The classification and regression results output by the detection head module are post-processed to generate the final ship target identification result. The post-processing includes confidence filtering and non-maximum suppression to remove duplicate detection boxes and low-confidence prediction results.

[0058] After post-processing, the obtained ship target recognition results include: category information of each ship target, used to characterize the specific ship type to which the detected ship target belongs; and location information of each ship target, used to characterize the specific location of the ship target in the standard ship image, wherein the location information is given in the form of bounding box coordinates.

[0059] Furthermore, the ship target detection network includes a YOLOv10 network and an LSDECD detection head, among which, The LSDECD detection head includes multiple input convolutional layers, a first detail enhancement convolutional layer, a second detail enhancement convolutional layer, multiple regression and classification convolutional groups, and multiple scaling layers. The input convolutional layer is used to receive ship feature maps at different scales extracted by the YOLOv10 backbone network; The first detail enhancement convolutional layer is used to fuse feature maps of different scales to generate a preliminary enhanced first multi-scale fusion feature, and then transmits the first multi-scale fusion feature to the second detail enhancement convolutional layer. The second detail enhancement convolutional layer is used to receive the first multi-scale fusion feature and generate a fully enhanced second multi-scale fusion feature, and distribute the second multi-scale fusion feature to multiple parallel regression and classification convolutional groups, and the second detail enhancement convolutional layer has the same structure as the first detail enhancement convolutional layer. Multiple regression and classification convolutional groups correspond one-to-one with different detection scales. Each regression and classification convolutional group includes a parallel regression convolutional branch and a classification convolutional branch. The regression convolutional branch is used to perform convolution processing on the second multi-scale fusion features to predict the bounding box coordinate parameters of the ship target. The classification convolutional branch is used to perform convolution processing on the second multi-scale fusion features to predict the class probability distribution of the ship target. Multiple scaling layers are connected to the output of each regression convolution branch. Each scaling layer is used to adaptively adjust and normalize the bounding box prediction results at the corresponding scale.

[0060] Furthermore, the first detail-enhancing convolutional layer includes a fifth convolutional layer, a central difference convolutional layer, an angular difference convolutional layer, a vertical difference convolutional layer, and a horizontal difference convolutional layer, wherein, The fifth convolutional layer is used to receive feature maps of different scales and form baseline features corresponding to the ship feature map; The central difference convolutional layer is used to calculate the central difference value between each pixel in the ship feature map and its surrounding neighboring pixels in order to extract radial gradient features centered on the pixel. The angular difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its diagonal neighboring pixels in order to extract gradient features in the diagonal direction. Vertical difference convolutional layers are used to calculate the difference between each pixel in the ship feature map and its vertically adjacent pixels to extract gradient features in the vertical direction. The horizontal difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its left and right horizontal neighboring pixels in order to extract the gradient features in the horizontal direction.

[0061] In this embodiment, the end-to-end defogging network, the super-resolution reconstruction network, and the ship target detection network are cascaded in a serial manner.

[0062] In one example, in a standard LSCD detection header, the GN_Conv1×1 module represents a group normalization (GroupNorm) + convolutional (Conv) operation, where 1×1 indicates a 1×1 kernel size. The two light green group normalization (GN_Conv3×3) convolutional modules share weights, the three regression (Conv_reg) convolutional modules share weights, and the three classification (Conv_cls) convolutional modules share weights. Each Conv_Box module is followed by a Scale module, which is primarily a scaling factor used to match the detection of targets at different scales.

[0063] Please see Figure 5 The GN_Conv1×1 module in the LSCD detector head is replaced with Enhanced Detail Convolution (DEConv). DEConv consists of five convolutional layers: one original convolutional layer and four parallel differential convolutional (DC) layers for feature extraction. The original convolutional layer captures intensity-level information, while the differential convolutional layers focus on enhancing gradient-level information. Differential convolutions compute pixel differences and use kernel weights to generate feature maps, thereby improving the expressiveness and generalization ability of the convolution. Typical variants of differential convolution include Central Differential Convolution (CDC) and Corner Differential Convolution (ADC), which utilize rearranged kernel weights to reduce computational cost. Vertical Differential Convolution (VDC) and Horizontal Differential Convolution (HDC) incorporate traditional edge operators into the convolutional layers, encoding gradient prior knowledge to learn valuable gradient information, thereby improving the performance and efficiency of the convolution.

[0064] DEConv exhibits unique properties in convolution operations. When multiple kernels of the same size operate on the same input with the same stride and padding, summing their outputs produces the same result as summing the kernels at the corresponding positions to produce the final output. This property ensures that parameter counting or inference time does not increase, making DEConv suitable for lightweight applications. The formula for calculating DEConv is as follows:

[0065] in, This represents the output feature map of the detail-enhancing convolutional layer. Representing augmented detail convolution operations, This represents the input ship feature map. (i=1:5) represents five convolutional kernels. This represents the convolution operation. This represents the transformation kernel after merging parallel convolutions. DEConv can effectively extract detailed features such as texture, shape, and temperature gradients, which perfectly meets the accuracy requirements of detection tasks. It significantly improves the network's ability to perceive and extract fine-grained features. Comparative experiments verify that the YOLOv10 model based on LSDECD (LSDECD-YOLOv10) achieves a higher mAP value on the FLSRD dataset, especially showing better accuracy and robustness when detecting small-sized, blurry-detailed ship targets.

[0066] Please see Figure 6 The three feature maps P3, P4, and P5 output from the neck are received by the various branches of the LSDECD detection head. For each branch, a 1×1 convolution operation with combined group normalization is first applied to unify the feature channel dimensions. Subsequently, a 3×3DEConv with two shared parameters is used to expand the receptive field, aggregating rich spatial context and multi-scale information while reducing the number of model parameters. Finally, the loss for decoupling the classification and regression tasks is calculated separately using standard 1×1 convolutions. In addition, a Scale module is introduced after each regression prediction branch. This module contains a dynamically adjustable coefficient φ to adaptively adjust the scale representation of the target object, thereby addressing the issue of differences in the target size range handled by different detection heads.

[0067] In this embodiment, by downsampling the original marine vessel images and combining them with an atmospheric scattering model to simulate fog, a dataset of foggy images with different fog concentrations and imaging conditions can be generated while maintaining the consistency of the vessel's structure and semantic information. This improves the algorithm's adaptability to complex sea conditions in real foggy weather and alleviates the problem of difficulty in collecting samples in real foggy weather. Furthermore, by using an end-to-end dehazing network containing instance normalization layers, embedded residual layers, and depthwise separable convolutional layers, multi-scale feature extraction and fusion are performed on the initial foggy images. This effectively weakens the influence of fog scattering, enhances image contrast and clarity, and suppresses noise amplification, while better preserving vessel edges and texture details. The dehazed original vessel images are then input into a super-resolution reconstruction network, where an attention mechanism is used to focus on enhancing the vessel target area. This helps to highlight key structural information of the vessel, suppress interference from the background sea surface and sky, and achieve high-quality reconstruction of distant, small-sized vessel targets, improving the overall resolution and target recognizability of the image. At the same time, by performing target detection on standard vessel images after dehazing and super-resolution reconstruction, the detection network can work under clearer and higher resolution input conditions, thereby significantly improving the accuracy and stability of vessel target detection and recognition.

[0068] Based on the above method, this application discloses a foggy long-range ship target recognition system based on multi-dimensional image cascade processing, referencing... Figure 7The foggy long-range ship target identification system 1 includes a data acquisition module 11, a data processing module 12, and a target identification module 13, wherein... The data acquisition module 11 is used to acquire multiple sets of marine ship images as the original dataset, perform downsampling on the original dataset to generate fog-free ship images, and perform fog simulation on the fog-free ship images based on the atmospheric scattering model to obtain the initial foggy images. The data processing module 12 is used to input the initial foggy image into the end-to-end defogging network, and extract the multi-scale features corresponding to the initial foggy image through the instance normalization layer, the embedded residual layer and the depth separable convolutional layer in the end-to-end defogging network, so that the end-to-end defogging network outputs the defogging original ship image. The original ship image is then input into the super-resolution reconstruction network, and the feature representation of the ship target region in the original ship image is enhanced according to the attention mechanism of the super-resolution reconstruction network, so as to reconstruct the standard ship image. The target recognition module 13 is used to input standard ship images into the ship target detection network, detect ship targets in the standard ship images, and enable the ship target detection network to output ship target recognition results, wherein the ship target recognition results include the category information and location information of each ship target.

[0069] In one example, the end-to-end dehazing network comprises sequentially connected instance normalization layers, embedding residual layers, and depthwise separable convolutional layers, wherein... The instance normalization layer is used to receive the initial foggy image and normalize the features in the initial foggy image; The embedding residual layer and the instance normalization layer are connected by multiple convolutional layers. The embedding residual layer includes a first embedding residual block and a second embedding residual block. The first embedding residual block is used to receive the initial foggy feature map after preliminary convolution processing and extract the low-level texture features and edge information of the ship target to obtain the first foggy feature map. The second embedding residual block is used to receive the first foggy feature map and extract the high-level semantic features and complex fog and haze pattern features of the ship target to obtain the second foggy feature map. The embedded residual layer and the depthwise separable convolutional layer are connected by a first connecting layer. The depthwise separable convolutional layer includes a first depthwise separable convolution, a second connecting layer, a third connecting layer, and a standard convolutional layer. The first depthwise separable convolution is used to receive the second foggy feature map and extract the spatial context information in the second foggy feature map through a combination of depthwise convolution and pointwise convolution. The second depthwise separable convolution is used to perform multi-scale feature extraction and optimization on the features fused by the second connecting layer. The standard convolutional layer is used to receive the fused features output by the third connecting layer. Through traditional convolution operations, the multidimensional features are mapped to the defogging parameter space to generate the original ship image after defogging.

[0070] In one example, the super-resolution reconstruction network includes a first convolutional layer, a channel attention mechanism module, a spatial attention mechanism module, a second convolutional layer, a first activation function layer, a third convolutional layer, and a fusion unit, wherein... The first convolutional layer is used to perform an initial convolutional transformation on the received original ship image to extract shallow features of the original ship image and transmit the shallow features to the channel attention mechanism module. The channel attention mechanism module is used to receive the shallow features output by the first convolutional layer, generate channel features and channel attention weight vectors corresponding to the channel features, and transmit the weighted channel features to the spatial attention mechanism module. The spatial attention mechanism module is used to receive weighted channel features, generate a spatial attention weight map and spatial attention features based on the weighted channel features, and transmit the spatial attention features to the fusion unit; The second convolutional layer is used to receive the original ship image, perform convolution processing on the original ship image to extract the mid-level structural features, and transmit the mid-level structural features to the first activation function layer and the fusion unit. The first activation function layer is used to perform nonlinear transformation on the mid-level structural features output by the second convolutional layer, and then transmits the nonlinear features obtained after the nonlinear transformation to the third convolutional layer. The third convolutional layer is used to extract nonlinear features through convolution to obtain residual features, and the residual features, spatial attention features, and original ship images are reconstructed in the fusion unit to obtain standard ship images.

[0071] In one example, the spatial attention mechanism module includes an average pooling layer, a max pooling layer, a fourth connection layer, a fourth convolutional layer, and a second activation function layer, wherein... The average pooling layer is used to receive the weighted channel features output by the channel attention mechanism module, and to perform global average pooling on the weighted channel features in the channel dimension to obtain a spatial statistical feature map. The max pooling layer is used to receive the weighted channel features output by the channel attention mechanism module, and to perform global max pooling operation on the weighted channel features in the channel dimension to obtain a spatial saliency feature map. The fourth connection layer is used to concatenate the spatial statistical feature map and the spatial saliency feature map along the channel dimension to generate a spatial fusion feature map containing the two channels. The fourth convolutional layer is used to receive the spatial fusion feature map, and performs spatial convolution processing on the spatial fusion feature map through convolution operation to generate spatial attention features; The second activation function layer is used to perform non-linear activation on the spatial attention features and normalize the spatial attention features to generate a spatial attention weight map.

[0072] In one example, the ship target detection network includes a YOLOv10 network and an LSDECD detection head, wherein, The LSDECD detection head includes multiple input convolutional layers, a first detail enhancement convolutional layer, a second detail enhancement convolutional layer, multiple regression and classification convolutional groups, and multiple scaling layers. The input convolutional layer is used to receive ship feature maps at different scales extracted by the YOLOv10 backbone network; The first detail enhancement convolutional layer is used to fuse feature maps of different scales to generate a preliminary enhanced first multi-scale fusion feature, and then transmits the first multi-scale fusion feature to the second detail enhancement convolutional layer. The second detail-enhancing convolutional layer is used to receive the first multi-scale fusion feature, generate a fully enhanced second multi-scale fusion feature, and distribute the second multi-scale fusion feature to multiple parallel regression and classification convolutional groups; Multiple regression and classification convolutional groups correspond one-to-one with different detection scales. Each regression and classification convolutional group includes a parallel regression convolutional branch and a classification convolutional branch. The regression convolutional branch is used to perform convolution processing on the second multi-scale fusion features to predict the bounding box coordinate parameters of the ship target. The classification convolutional branch is used to perform convolution processing on the second multi-scale fusion features to predict the class probability distribution of the ship target. Multiple scaling layers are connected to the output of each regression convolution branch. Each scaling layer is used to adaptively adjust and normalize the bounding box prediction results at the corresponding scale.

[0073] In one example, the first detail enhancement convolutional layer includes a fifth convolutional layer, a central difference convolutional layer, an angular difference convolutional layer, a vertical difference convolutional layer, and a horizontal difference convolutional layer, wherein, The fifth convolutional layer is used to receive feature maps of different scales and form baseline features corresponding to the ship feature map; The central difference convolutional layer is used to calculate the central difference value between each pixel in the ship feature map and its surrounding neighboring pixels in order to extract radial gradient features centered on the pixel. The angular difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its diagonal neighboring pixels in order to extract gradient features in the diagonal direction. Vertical difference convolutional layers are used to calculate the difference between each pixel in the ship feature map and its vertically adjacent pixels to extract gradient features in the vertical direction. The horizontal difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its left and right horizontal neighboring pixels in order to extract the gradient features in the horizontal direction.

[0074] In one example, the end-to-end defogging network, the super-resolution reconstruction network, and the ship target detection network are cascaded in a serial manner.

[0075] Please see Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 2 may include: at least one processor 21, at least one network interface 24, user interface 23, memory 25, and at least one communication bus 22.

[0076] The communication bus 22 is used to enable communication between these components.

[0077] The user interface 23 may include a display screen and a camera. Optionally, the user interface 23 may also include a standard wired interface and a wireless interface.

[0078] The network interface 24 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0079] The processor 21 may include one or more processing cores. The processor 21 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 25, and by calling data stored in the memory 25. Optionally, the processor 21 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 21 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 21.

[0080] The memory 25 may include random access memory (RAM) or read-only memory. Optionally, the memory 25 may include non-transitory computer-readable storage medium. The memory 25 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 25 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 25 may also be at least one storage device located remotely from the aforementioned processor 21. Figure 8 As shown, the memory 25, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a foggy long-distance ship target recognition method based on multi-dimensional image cascade processing.

[0081] exist Figure 8 In the electronic device 2 shown, the user interface 23 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 21 can be used to call the application program stored in the memory 25, which is a method for identifying distant ships in foggy weather based on multi-dimensional image cascade processing. When executed by one or more processors, the electronic device performs one or more methods as described in the above embodiments.

[0082] A non-transitory computer-readable storage medium stores instructions that, when executed by one or more processors, cause a computer to perform one or more methods as described in the above embodiments.

[0083] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0084] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0085] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0086] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0087] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying a long-distance ship target in foggy weather based on multi-dimensional image cascade processing, characterized in that, The method includes: Multiple sets of images of ships at sea are collected as the original dataset. The original dataset is downsampled to generate fog-free ship images. Fog simulation is performed on the fog-free ship images based on an atmospheric scattering model to obtain initial foggy images. An initial foggy image is input into an end-to-end defogging network, and multi-scale features corresponding to the initial foggy image are extracted through the instance normalization layer, the embedded residual layer and the depth-separable convolutional layer in the end-to-end defogging network, so that the end-to-end defogging network outputs the original ship image after defogging. The original ship image is input into a super-resolution reconstruction network, and the feature representation of the ship target region in the original ship image is enhanced according to the attention mechanism of the super-resolution reconstruction network to reconstruct a standard ship image. The standard ship image is input into the ship target detection network to detect ship targets in the standard ship image, so that the ship target detection network outputs ship target recognition results, wherein the ship target recognition results include the category information and location information of each ship target.

2. The method for identifying a ship target at a long distance in foggy weather based on multi-dimensional image cascade processing according to claim 1, characterized in that, The end-to-end dehazing network comprises, in sequence, an instance normalization layer, an embedded residual layer, and a depthwise separable convolutional layer, wherein, The instance normalization layer is used to receive the initial foggy image and normalize the features in the initial foggy image. The embedding residual layer and the instance normalization layer are connected through multiple convolutional layers. The embedding residual layer includes a first embedding residual block and a second embedding residual block. The first embedding residual block is used to receive the initial foggy feature map after preliminary convolution processing and extract the low-level texture features and edge information of the ship target to obtain the first foggy feature map. The second embedding residual block is used to receive the first foggy feature map and extract the high-level semantic features and complex fog and haze pattern features of the ship target to obtain the second foggy feature map. The embedded residual layer and the depthwise separable convolutional layer are connected by a first connection layer. The depthwise separable convolutional layer includes a first depthwise separable convolution, a second connection layer, a third connection layer, and a standard convolutional layer. The first depthwise separable convolution is used to receive the second foggy feature map and extract the spatial context information in the second foggy feature map through a combination of depthwise convolution and pointwise convolution. The second depthwise separable convolution is used to perform multi-scale feature extraction and optimization on the features fused by the second connection layer. The standard convolutional layer is used to receive the fused features output by the third connection layer and map the multi-dimensional features to the defogging parameter space through traditional convolution operations to generate the original defogging ship image.

3. The foggy long-distance ship target recognition method based on multi-dimensional image cascade processing as described in claim 1, characterized in that, The super-resolution reconstruction network includes a first convolutional layer, a channel attention mechanism module, a spatial attention mechanism module, a second convolutional layer, a first activation function layer, a third convolutional layer, and a fusion unit, wherein... The first convolutional layer is used to perform an initial convolutional transformation on the received original ship image to extract shallow features of the original ship image and transmit the shallow features to the channel attention mechanism module; The channel attention mechanism module is used to receive the shallow features output by the first convolutional layer, generate channel features and channel attention weight vectors corresponding to the channel features, and transmit the weighted channel features to the spatial attention mechanism module. The spatial attention mechanism module is used to receive the weighted channel features, generate a spatial attention weight map and spatial attention features based on the weighted channel features, and transmit the spatial attention features to the fusion unit; The second convolutional layer is used to receive the original ship image, perform convolution processing on the original ship image to extract mid-level structural features, and transmit the mid-level structural features to the first activation function layer and the fusion unit; The first activation function layer is used to perform a nonlinear transformation on the mid-level structural features output by the second convolutional layer, and then transmits the nonlinear features obtained after the nonlinear transformation to the third convolutional layer. The third convolutional layer is used to extract the nonlinear features through convolution to obtain residual features, and the residual features, the spatial attention features, and the original ship image are reconstructed in the fusion unit to obtain a standard ship image.

4. The foggy long-distance ship target recognition method based on multi-dimensional image cascade processing as described in claim 3, characterized in that, The spatial attention mechanism module includes an average pooling layer, a max pooling layer, a fourth connection layer, a fourth convolutional layer, and a second activation function layer, wherein... The average pooling layer is used to receive the weighted channel features output by the channel attention mechanism module, and to perform a global average pooling operation on the weighted channel features in the channel dimension to obtain a spatial statistical feature map. The max pooling layer is used to receive the weighted channel features output by the channel attention mechanism module, and to perform a global max pooling operation on the weighted channel features in the channel dimension to obtain a spatial saliency feature map. The fourth connection layer is used to concatenate the spatial statistical feature map and the spatial saliency feature map in the channel dimension to generate a spatial fusion feature map containing two channels. The fourth convolutional layer is used to receive the spatial fusion feature map and perform spatial convolution processing on the spatial fusion feature map through convolution operations to generate spatial attention features; The second activation function layer is used to perform nonlinear activation on the spatial attention features and normalize the spatial attention features to generate a spatial attention weight map.

5. The foggy long-distance ship target recognition method based on multi-dimensional image cascade processing as described in claim 1, characterized in that, The ship target detection network includes a YOLOv10 network and an LSDECD detection head, wherein... The LSDECD detection head includes multiple input convolutional layers, a first detail enhancement convolutional layer, a second detail enhancement convolutional layer, multiple regression and classification convolutional groups, and multiple scaling layers, wherein... The input convolutional layer is used to receive ship feature maps of different scales extracted by the YOLOv10 backbone network; The first detail enhancement convolutional layer is used to fuse feature maps of different scales to generate a first multi-scale fusion feature with preliminary enhancement, and then transmits the first multi-scale fusion feature to the second detail enhancement convolutional layer. The second detail-enhancing convolutional layer is used to receive the first multi-scale fusion feature, generate a sufficiently enhanced second multi-scale fusion feature, and distribute the second multi-scale fusion feature to multiple parallel regression-classification convolutional groups; The multiple regression and classification convolutional groups correspond one-to-one with different detection scales. Each regression and classification convolutional group includes a parallel regression convolutional branch and a classification convolutional branch. The regression convolutional branch is used to perform convolution processing on the second multi-scale fusion feature to predict the bounding box coordinate parameters of the ship target. The classification convolutional branch is used to perform convolution processing on the second multi-scale fusion feature to predict the class probability distribution of the ship target. The multiple scaling layers are connected to the output of each regression convolution branch, and each scaling layer is used to perform scale adaptive adjustment and normalization on the bounding box prediction results of the corresponding scale.

6. The foggy long-distance ship target recognition method based on multi-dimensional image cascade processing as described in claim 5, characterized in that, The first detail-enhancing convolutional layer includes a fifth convolutional layer, a central difference convolutional layer, an angular difference convolutional layer, a vertical difference convolutional layer, and a horizontal difference convolutional layer, wherein, The fifth convolutional layer is used to receive feature maps of different scales and form a reference feature corresponding to the ship feature map; The central difference convolutional layer is used to calculate the central difference value between each pixel in the ship feature map and its surrounding neighboring pixels, so as to extract radial gradient features centered on the pixel. The angular difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its diagonal neighboring pixels in order to extract gradient features in the diagonal direction. The vertical difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its vertically adjacent pixels above and below it, so as to extract the gradient features in the vertical direction. The horizontal difference convolutional layer is used to calculate the difference between each pixel in the ship feature map and its horizontal left and right neighboring pixels to extract the gradient features in the horizontal direction.

7. The foggy long-distance ship target recognition method based on multi-dimensional image cascade processing as described in claim 1, characterized in that, The end-to-end defogging network, the super-resolution reconstruction network, and the ship target detection network are cascaded in a serial manner.

8. A foggy long-range ship target recognition system based on multi-dimensional image cascade processing, characterized in that, The foggy long-range ship target identification system (1) includes a data acquisition module (11), a data processing module (12), and a target identification module (13), wherein, The data acquisition module (11) is used to acquire multiple sets of marine ship images as the original dataset, perform downsampling on the original dataset to generate fog-free ship images, and perform fog simulation on the fog-free ship images based on the atmospheric scattering model to obtain the initial foggy image; The data processing module (12) is used to input the initial foggy image into the end-to-end defogging network, and extract the multi-scale features corresponding to the initial foggy image through the instance normalization layer, the embedded residual layer and the depth separable convolutional layer in the end-to-end defogging network, so that the end-to-end defogging network outputs the defogging original ship image, inputs the original ship image into the super-resolution reconstruction network, and enhances the feature expression of the ship target region in the original ship image according to the attention mechanism of the super-resolution reconstruction network, so as to reconstruct the standard ship image; The target recognition module (13) is used to input the standard ship image into the ship target detection network and detect the ship targets in the standard ship image so that the ship target detection network outputs the ship target recognition result, wherein the ship target recognition result includes the category information and location information of each ship target.

9. An electronic device, characterized in that, The device includes a processor (21), a memory (25), a user interface (23), and a network interface (24), wherein the memory (25) is used to store instructions, the user interface (23) and the network interface (24) are used to communicate with other devices, and the processor (21) is used to execute the instructions stored in the memory (25) to cause the electronic device (2) to perform the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images