Sea surface target identification method and system

Abstract

The invention discloses a sea surface target identification method and system, and the method comprises the steps: obtaining a plurality of frames of target images of a sea surface live-action video, carrying out the preprocessing of each frame of target image, and obtaining a plurality of frames of preprocessed target images; performing image data optimization on each frame of preprocessed target image by adopting a data enhancement method based on hybrid enhancement to obtain a plurality of frames of optimized target images; slicing the plurality of frames of optimized target images to obtain a plurality of frames of slice images; constructing an improved YOLOv8 model, and performing target detection on the plurality of frames of slice images based on the improved YOLOv8 model to obtain a corresponding target detection result; and fusing the target detection results of the plurality of frames of slice images to obtain an overall target detection result. According to the invention, the algorithm complexity can be effectively reduced while the high performance is achieved, and the detection of offshore and offshore targets can be realized with low cost.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method and system for identifying sea surface targets. Background Art

[0002] Sea surface target recognition technology is of great significance in the fields of ocean monitoring and maritime search and rescue. With the development of computer vision and deep learning technology, automated and intelligent target recognition technology has received widespread attention and application. Traditional sea surface target recognition methods based on radar detection have the ability of all-weather and long-distance detection and can work under severe weather conditions. However, such methods are limited by the influence of radar band frequency. The resolution of radar is relatively low, making it difficult to accurately identify and classify targets. At the same time, due to the large attenuation of electromagnetic waves in water and the susceptibility to interference from factors such as sea surface waves, the detection accuracy is reduced. In the face of the task of sea surface target recognition, radar detection methods are generally more suitable for ultra-long-range radar warnings, using land-based or air-based radars to achieve radar warnings. However, for the task of precise target detection, identification and classification in the offshore area, radar detection-based methods are insufficient in resolution and detection performance. The traditional sonar technology used for underwater detection has excellent performance in underwater detection and can detect diving targets and underwater environments. However, the range of sonar detection is limited and is greatly affected by underwater noise interference. In addition, for the detection of surface targets, the sonar-based detection method cannot achieve accurate target recognition and classification, especially for static targets. Sonar detection cannot achieve target recognition well, and the layout conditions of sonar equipment are high and the equipment is expensive. In the field of sea surface target recognition, there are problems such as high deployment and maintenance costs and low detection performance. In view of the shortcomings of the above-mentioned traditional surface target recognition algorithms, with the rapid development of artificial intelligence technology, image recognition methods based on deep learning have gradually become an important research direction for sea surface target recognition. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art. The present invention provides a method and system for identifying sea surface targets, which can effectively reduce algorithm complexity while achieving high performance, and can realize the detection of near-shore and offshore targets at a lower cost and price.

[0004] In order to solve the above technical problems, the present invention provides a method for identifying sea surface targets, the method comprising:

[0005] Acquire several frames of target images of the sea surface real scene video, and preprocess each frame of the target image to obtain several frames of preprocessed target images;

[0006] The data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image to obtain several frames of optimized target images;

[0007] Slicing a plurality of frames of optimized target images to obtain a plurality of frames of sliced ​​images;

[0008] Construct an improved YOLOv8 model, perform target detection on several frames of slice images based on the improved YOLOv8 model, and obtain corresponding target detection results;

[0009] The target detection results of several frames of slice images are fused to obtain the overall target detection result.

[0010] Optionally, preprocessing each frame of the target image to obtain a plurality of frames of preprocessed target images includes:

[0011] Perform denoising on each frame of the target image to obtain several frames of denoised target images;

[0012] Image enhancement processing is performed on each frame of the target image after denoising to obtain several frames of target images after preprocessing.

[0013] Optionally, the data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image to obtain a plurality of frames of optimized target images, including:

[0014] Perform segmentation processing on each frame of the preprocessed target image to obtain several frames of segmented images;

[0015] Performing preliminary image enhancement processing on a plurality of frames of segmented images to obtain a plurality of frames of preliminary enhanced segmented images;

[0016] Linearly superimposing several frames of preliminary enhanced segmented images to obtain preliminary enhanced segmented images after linear superposition of several frames;

[0017] Based on the mosaic principle and Cutout method, the preliminary enhanced segmented images after linear superposition of several frames are spliced ​​and optimized to obtain several frames of optimized target images.

[0018] Optionally, slicing a plurality of frames of optimized target images to obtain a plurality of frames of sliced ​​images includes:

[0019] Based on the sliding window, several frames of optimized target images are segmented into regions to obtain several frames of sliced ​​images.

[0020] Optionally, the structure of the improved YOLOv8 model includes a backbone network, a neck network and a detection head.

[0021] Optionally, the backbone network adopts a cross-stage local network, the cross-stage local network includes several convolutional layers and a feature pyramid network, the neck network includes a feature pyramid network, a path aggregation network and an upsampling layer, and the detection head includes a small target detection layer.

[0022] Optionally, the feature pyramid network is used to convey strong semantic features from top to bottom, the path aggregation network is used to convey positioning features from bottom to top, and the detection head is used to perform target detection prediction at different feature levels.

[0023] Optionally, several frames of sliced ​​images generated by optimizing and slicing several frames of pre-processed target images by a data enhancement method based on hybrid enhancement are used as a training data set, and the training data set is input into the improved YOLOv8 model for training.

[0024] Optionally, fusing the target detection results of several frames of slice images to obtain an overall target detection result includes:

[0025] In the process of fusing the target detection results of several frame slice images, the target detection results of all frame slice images are subjected to overlapping detection frame removal processing based on the non-maximum suppression method to obtain several target detection frames, and the overall target detection result is generated based on the several target detection frames.

[0026] In addition, the present invention also provides a sea surface target recognition system, the system comprising:

[0027] Image acquisition and preprocessing module: used to acquire several frames of target images of the sea surface real scene video, and preprocess each frame of the target image to obtain several frames of preprocessed target images;

[0028] Image data optimization module: used to optimize the image data of each frame of the pre-processed target image by using a data enhancement method based on hybrid enhancement, and obtain several frames of optimized target images;

[0029] Image slicing module: used to slice several frames of optimized target images to obtain several frames of sliced ​​images;

[0030] Target detection module: used to build an improved YOLOv8 model, perform target detection on several frame slice images based on the improved YOLOv8 model, and obtain corresponding target detection results;

[0031] Target detection fusion module: used to fuse the target detection results of several frame slice images to obtain the overall target detection result.

[0032] In the embodiment of the present invention, a data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image, and a high-performance data enhancement method is used to solve the difficulties such as multi-target overlap. An improved YOLOv8 model is constructed, and a small target detection layer is set to detect small targets on the sea surface. Target detection is performed on several frames of slice images based on the improved YOLOv8 model to obtain corresponding target detection results, and the target detection results of several frames of slice images are fused to achieve slice hyper-reasoning based on the required target location. While achieving high performance, the algorithm complexity can be effectively reduced, and the detection of offshore and coastal targets can be achieved at a lower cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 is a schematic diagram of a flow chart of a method for identifying a sea surface target in an embodiment of the present invention;

[0035] Figure 2 It is a schematic diagram of the structure of the sea surface target recognition system in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] Embodiment 1

[0038] See also Figure 1 , Figure 1 : is a flow chart of a method for identifying a sea surface target in an embodiment of the present invention, the method comprising:

[0039] S11: acquiring several frames of target images of the sea surface real scene video, and preprocessing each frame of the target image to obtain several frames of preprocessed target images;

[0040] In the specific implementation process of the present invention, the preprocessing of each frame of the target image to obtain several frames of preprocessed target images includes: performing denoising processing on each frame of the target image to obtain several frames of denoised target images; performing image enhancement processing on each frame of the denoised target image to obtain several frames of preprocessed target images.

[0041] Specifically, several frames of target images of the sea surface real scene video are obtained through the acquisition device. Each frame of the target image is denoised, and each frame of the target image is preliminarily denoised using the two-dimensional discrete cosine transform domain to obtain several frames of target images after the preliminary denoising, and the energy parameters of several frames of target images after the preliminary denoising are obtained. According to the energy parameters, the residual amount of the central pixel of each unit area of ​​several frames of target images after the preliminary denoising is obtained, and the weight matrix of each unit area is calculated based on the residual amount. Based on the weight matrix, several frames of target images after the preliminary denoising are further denoised to obtain several frames of target images after the preliminary denoising. Image enhancement is performed on each frame of the target image after the denoising, and the object field of each frame of the target image after the denoising is modulated in the frequency domain, and the frequency domain is Fourier transformed, and the imaginary part of the frequency domain after the Fourier transform is taken, and the absolute value of the imaginary part is calculated, and the power value is calculated based on the absolute value of the imaginary part after summing along the frequency, and the image is enhanced based on the power value. After the image enhancement is completed, the image preprocessing is completed, and several frames of preprocessed target images are obtained.

[0042] S12: optimizing the image data of each frame of the preprocessed target image by using a data enhancement method based on hybrid enhancement to obtain a plurality of frames of optimized target images;

[0043] In the specific implementation process of the present invention, the data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image to obtain several frames of optimized target images, including: segmenting each frame of the preprocessed target image to obtain several frames of segmented images; performing preliminary image enhancement processing on several frames of segmented images to obtain several frames of preliminary enhanced segmented images; linearly superimposing several frames of preliminary enhanced segmented images to obtain several frames of preliminary enhanced segmented images after linear superposition; splicing and optimizing several frames of preliminary enhanced segmented images after linear superposition based on the mosaic principle and the Cutout method to obtain several frames of optimized target images.

[0044] Specifically, segmentation is performed on each frame of the preprocessed target image, and a neural network is used to assist the segmentation of each frame of the preprocessed target image to obtain several frames of segmented images. Then, the several frames of segmented images are classified and screened, and the segmentation errors are manually corrected. Objects with obvious features and completeness are selected to form a database for subsequent data enhancement processing. Preliminary image enhancement is performed on several frames of segmented images, and image changes such as random cropping, rotation, and flipping are performed on several frames of segmented images to obtain several frames of preliminary enhanced segmented images. Several frames of preliminary enhanced segmented images are linearly superimposed, that is, every two frames of preliminary enhanced segmented images are linearly superimposed at a random ratio, and the labels are adjusted accordingly to obtain several frames of preliminary enhanced segmented images after linear superposition. Based on the mosaic principle and the Cutout method, the preliminary enhanced segmented images after linear superposition of several frames are spliced ​​and optimized. The mosaic principle is an image processing technology that is often used to hide or protect sensitive information. Its principle is to divide the specific area of ​​the image into small blocks and fill each small block with the same color or blur effect, divide the initial enhanced segmentation image after linear superposition of each frame into multiple small blocks, and the divided small blocks are of the same size. Based on the blur filter, the recognition of each small block is removed to make it lose its recognition, and the small blocks after the recognition removal are recombined into a mosaic image, so as to achieve the purpose of hiding or confusing the target area. Cutout is a data enhancement technology, which is achieved by masking random parts of the input image, randomly covering parts of the image, simulating occlusion to enhance the situation of occlusion of sea surface targets, and enhancing the recognition ability of the model when the target is partially occluded, and obtaining several frames of optimized target images.

[0045] S13: Slicing a plurality of frames of optimized target images to obtain a plurality of frames of sliced ​​images;

[0046] In the specific implementation process of the present invention, slicing the plurality of frames of optimized target images to obtain the plurality of frames of sliced ​​images includes: performing region segmentation on the plurality of frames of optimized target images based on a sliding window to obtain the plurality of frames of sliced ​​images.

[0047] Specifically, the window size and step size of the sliding window are set, and then the sliding window is set. Based on the sliding window, several frames of optimized target images are segmented to obtain several frames of sliced ​​images. The image slicing technology ensures that each slice contains enough detail information to avoid small targets being ignored in large-resolution images. At the same time, the slice super-inference method decomposes the large-resolution image into several small blocks, reducing the amount of calculation for a single inference and improving the inference efficiency.

[0048] S14: construct an improved YOLOv8 model, and perform target detection on a plurality of frame slice images based on the improved YOLOv8 model to obtain corresponding target detection results;

[0049] In the specific implementation process of the present invention, the structure of the improved YOLOv8 model includes a backbone network, a neck network and a detection head.

[0050] Furthermore, the backbone network adopts a cross-stage local network, which includes several convolutional layers and a feature pyramid network, the neck network includes a feature pyramid network, a path aggregation network and an upsampling layer, and the detection head includes a small target detection layer.

[0051] Furthermore, the feature pyramid network is used to convey strong semantic features from top to bottom, the path aggregation network is used to convey positioning features from bottom to top, and the detection head is used to perform target detection prediction at different feature levels.

[0052] Furthermore, several frames of sliced ​​images generated by optimizing and slicing several frames of preprocessed target images by a data enhancement method based on hybrid enhancement are used as training data sets, and the training data sets are input into the improved YOLOv8 model for training.

[0053] Specifically, facing the complexity of the marine environment, such as variable lighting conditions, wave interference and target morphological diversity, it is often difficult to obtain satisfactory recognition results by training models based solely on raw image data. In order to improve the generalization ability and robustness of the model, data enhancement technology is widely used in the training process of deep learning models. In response to the above problems, a high-performance data enhancement method based on hybrid enhancement is designed. First, the basic data enhancement method is used to generate more samples by performing various transformations and processing on the training data. Its main purpose is to increase the diversity of the data set, prevent model overfitting, and improve the generalization ability of the model in different environments. First, the network is trained, and after the training is completed, the neural network is used to assist in segmentation. Then, the segmentation results are classified and screened, and the segmentation errors are manually corrected. Objects with obvious features and completeness are selected to form a database for subsequent data enhancement processing. Another expansion method is to directly search for images of objects such as buoys, sailboats, speedboats, etc. in the search engine, and segment and save them using the neural network according to the above method for subsequent processing. After obtaining the segmented objects, the designed data enhancement method is to use a script to paste the segmented images obtained in the previous step into the data set according to the size and category as much as possible according to the characteristics of the original data set. For example: ships in the distance are smaller and closer to the coastline; ships in the near distance are larger and located below the coastline; the coast is above the coastline, etc. The above data enhancements are all performed randomly, so it is necessary to manually judge whether the generated images are reasonable in the end, and then put them into the training set for training. After that, find other international open source data sets, filter out pictures with obstacles such as ships and buoys, and add them to the training set for training. In the training set, a data enhancement method based on hybrid enhancement is used. First, traditional image changes such as random cropping, rotation, and flipping are used to increase the diversity of training samples and enhance the robustness of the model. Then, according to the particularity of the sea surface target recognition task, a hybrid enhancement method is used in the training process to linearly superimpose two images at random proportions and adjust the labels accordingly. Generating new training samples in this way helps the model better learn the relationship between different targets and helps improve the recognition performance in complex sea environments, especially for the case of multiple target overlaps. This method can effectively solve the target recognition problem in the case of multiple target confusion. At the same time, the Cutout method and the Mosaic principle are used to randomly cover parts of the image, simulate occlusion, and enhance the recognition ability of the model when the target is partially occluded. Since the YOLO model does not match the image scale in the problem to be solved, and in order to enhance the generalization ability of the model, the Mosaic principle is used to stitch together multiple images to enhance the detection ability of targets at different scales and increase the diversity of the data set.Adversarial training is also used in the training process. By generating adversarial samples to train the model, it can resist potential adversarial attacks and enhance the robustness of the model. In the recognition of sea surface targets, images with noise or other interference factors can be generated for training to improve the performance of the model in harsh environments. Through the above data enhancement method based on hybrid enhanced intelligence, the model can use adversarial training methods in more diverse data sets to achieve better performance.

[0054] The second problem for sea surface target recognition is that the YOLO (You Only Look Once) neural network has a poor recognition effect on small targets. The YOLO neural network is an object detection system for computer vision tasks. The YOLOv8 model is an object detection model based on the YOLO neural network. There are many small targets at long distances in the data set. First, the model architecture of YOLOv8 is improved. After analyzing the small targets in the data set, it is found that these small targets are caused by the long distance of the lens. Such targets are relatively complex, usually the side or front of the ship, and there are also some long-distance buoys, so it is difficult to recognize such targets. In response to this situation, a detection layer is added on the basis of YOLOv8 to improve the feature extraction ability of small targets. Since the sample size of small targets is small and the sampling multiple of YOLOv8 is relatively large, it is difficult for the model to learn the feature information of small targets for deeper target feature maps. Based on the original network model of YOLOv8, the feature extraction network is modified to achieve small target detection by adding upsampling layers and convolution layers. In addition, parameters such as the size of the convolution kernel are modified according to the target features to better recognize the distant sailboats in the data set. The YOLOv8 model uses a cross-stage local network as the backbone network, which consists of a series of convolution layers and a feature pyramid network. The output feature map of the backbone network will be passed to the subsequent processing layer. The network structure design of the neck network applies the structure of the feature pyramid network plus the path aggregation network. The feature pyramid network layer conveys strong semantic features from top to bottom, while the path aggregation network tower conveys positioning features from bottom to top. The detection head is responsible for predicting the bounding box, category and confidence of the target. Detection heads of different scales are responsible for target detection prediction at different feature levels. Compared with the original model, the input resolution of the backbone network is increased to 1056, the neck network part adds an upsampling layer, sets a new convolution kernel size for small targets, and the detection head part is changed from three layers to four layers, expanding the small target detection layer. In order to improve the generalization ability of the model and enhance the detection performance of the small target detection layer, the modified model is pre-trained on the Common Objects in Context (COCO) dataset, and the complexity of the model is reduced by model pruning. The pre-trained model is transferred to the dataset for training to obtain a trained improved YOLOv8 model. Several frames of slice images are input into the trained improved YOLOv8 model to obtain the target detection result corresponding to each frame of slice image.

[0055] S15: Fusing the target detection results of several frames of slice images to obtain an overall target detection result.

[0056] In the specific implementation process of the present invention, the target detection results of several frame slice images are fused to obtain the overall target detection result, including: in the process of fusing the target detection results of several frame slice images, the target detection results of all frame slice images are subjected to overlapping detection frame removal processing based on the non-maximum suppression method to obtain several target detection frames, and the overall target detection result is generated based on the several target detection frames.

[0057] Specifically, the target detection method mainly uses low-resolution images (640x480) on the COCO dataset. In these datasets, the targets are relatively large objects and have large pixel coverage, which covers 60% of the image height on average. The models pre-trained using these datasets provide very successful detection performance for similar inputs. However, the accuracy in small object detection tasks is significantly lower in high-resolution images. The image slicing assisted super reasoning method is to divide the image into several regions through a sliding window, predict each region separately, then merge the prediction results of each region with the prediction results of the entire image, and finally filter with non-maximum suppression. The pixel area of ​​small targets is relatively large relative to the image input to the network. For the analysis of the sea surface target recognition task, the image slicing assisted super reasoning method can effectively realize the recognition of small targets. In order to improve the detection performance of small targets at ultra-long distances, the image slicing assisted super reasoning method based on spatial attention is adopted. This method realizes efficient and accurate sea surface target recognition by introducing spatial attention mechanism and image slicing technology. By performing target detection on each slice image separately and introducing spatial attention mechanism, the model can better pay attention to the location of small targets and improve detection accuracy. Finally, the detection results of all slices are fused, and the non-maximum suppression method is used to remove the overlapping detection frames, that is, the slice detection results are reconstructed into the original image to form a complete target detection output. According to the confidence threshold of the detection frame, the low-confidence detection results are filtered and the high-confidence target detection frame is retained. After the introduction of the spatial attention mechanism, the model can pay more attention to the location of small targets more accurately, improving the detection accuracy of small targets. The image slicing technology ensures that each slice contains enough detail information to avoid small targets being ignored in large-resolution images. At the same time, the slice super-inference method decomposes the large-resolution image into several small blocks, reducing the computational amount of a single inference and improving the inference efficiency. The image slicing assisted super-inference method based on spatial attention effectively improves the performance of YOLOv8 in sea surface target recognition by introducing the spatial attention mechanism and image slicing technology. This method performs well in improving the detection accuracy of small targets, enhancing robustness in complex backgrounds, and improving computational efficiency. The image slicing super-inference method based on spatial attention has a very significant improvement in improving the recognition probability of small targets. However, for coastal targets, if the intersection-over-union ratio is set to a large value, non-maximum suppression sometimes does not work properly. For this reason, this method is only used when identifying obstacle labels, and the prediction results from the full image are used when identifying the coast. The fusion of the two can obtain the best result.

[0058] In the embodiment of the present invention, a data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image, and a high-performance data enhancement method is used to solve the difficulties such as multi-target overlap. An improved YOLOv8 model is constructed, and a small target detection layer is set to detect small targets on the sea surface. Target detection is performed on several frames of slice images based on the improved YOLOv8 model to obtain corresponding target detection results, and the target detection results of several frames of slice images are fused to achieve slice hyper-reasoning based on the required target location. While achieving high performance, the algorithm complexity can be effectively reduced, and the detection of offshore and coastal targets can be achieved at a lower cost.

[0059] Embodiment 2

[0060] See also Figure 2 , Figure 2 : is a schematic diagram of the structure of a sea surface target recognition system in an embodiment of the present invention, the system comprises:

[0061] The image acquisition and preprocessing module 21 is used to acquire several frames of target images of the sea surface real scene video, and preprocess each frame of the target image to obtain several frames of preprocessed target images;

[0062] In the specific implementation process of the present invention, the preprocessing of each frame of the target image to obtain several frames of preprocessed target images includes: performing denoising processing on each frame of the target image to obtain several frames of denoised target images; performing image enhancement processing on each frame of the denoised target image to obtain several frames of preprocessed target images.

[0063] Specifically, several frames of target images of the sea surface real scene video are obtained through the acquisition device. Each frame of the target image is denoised, and each frame of the target image is preliminarily denoised using the two-dimensional discrete cosine transform domain to obtain several frames of target images after the preliminary denoising, and the energy parameters of several frames of target images after the preliminary denoising are obtained. According to the energy parameters, the residual amount of the central pixel of each unit area of ​​several frames of target images after the preliminary denoising is obtained, and the weight matrix of each unit area is calculated based on the residual amount. Based on the weight matrix, several frames of target images after the preliminary denoising are further denoised to obtain several frames of target images after the preliminary denoising. Image enhancement is performed on each frame of the target image after the denoising, and the object field of each frame of the target image after the denoising is modulated in the frequency domain, and the frequency domain is Fourier transformed, and the imaginary part of the frequency domain after the Fourier transform is taken, and the absolute value of the imaginary part is calculated, and the power value is calculated based on the absolute value of the imaginary part after summing along the frequency, and the image is enhanced based on the power value. After the image enhancement is completed, the image preprocessing is completed, and several frames of preprocessed target images are obtained.

[0064] Image data optimization module 22: used to optimize the image data of each frame of the pre-processed target image by using a data enhancement method based on hybrid enhancement, so as to obtain a plurality of frames of optimized target images;

[0065] In the specific implementation process of the present invention, the data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image to obtain several frames of optimized target images, including: segmenting each frame of the preprocessed target image to obtain several frames of segmented images; performing preliminary image enhancement processing on several frames of segmented images to obtain several frames of preliminary enhanced segmented images; linearly superimposing several frames of preliminary enhanced segmented images to obtain several frames of preliminary enhanced segmented images after linear superposition; splicing and optimizing several frames of preliminary enhanced segmented images after linear superposition based on the mosaic principle and the Cutout method to obtain several frames of optimized target images.

[0066] Specifically, segmentation is performed on each frame of the preprocessed target image, and a neural network is used to assist the segmentation of each frame of the preprocessed target image to obtain several frames of segmented images. Then, the several frames of segmented images are classified and screened, and the segmentation errors are manually corrected. Objects with obvious features and completeness are selected to form a database for subsequent data enhancement processing. Preliminary image enhancement is performed on several frames of segmented images, and image changes such as random cropping, rotation, and flipping are performed on several frames of segmented images to obtain several frames of preliminary enhanced segmented images. Several frames of preliminary enhanced segmented images are linearly superimposed, that is, every two frames of preliminary enhanced segmented images are linearly superimposed at a random ratio, and the labels are adjusted accordingly to obtain several frames of preliminary enhanced segmented images after linear superposition. Based on the mosaic principle and the Cutout method, the preliminary enhanced segmented images after linear superposition of several frames are spliced ​​and optimized. The mosaic principle is an image processing technology that is often used to hide or protect sensitive information. Its principle is to divide the specific area of ​​the image into small blocks and fill each small block with the same color or blur effect, divide the initial enhanced segmentation image after linear superposition of each frame into multiple small blocks, and the divided small blocks are of the same size. Based on the blur filter, the recognition of each small block is removed to make it lose its recognition, and the small blocks after the recognition removal are recombined into a mosaic image, so as to achieve the purpose of hiding or confusing the target area. Cutout is a data enhancement technology, which is achieved by masking random parts of the input image, randomly covering parts of the image, simulating occlusion to enhance the situation of occlusion of sea surface targets, and enhancing the recognition ability of the model when the target is partially occluded, and obtaining several frames of optimized target images.

[0067] Image slicing module 23: used for slicing several frames of optimized target images to obtain several frames of sliced ​​images;

[0068] In the specific implementation process of the present invention, slicing the plurality of frames of optimized target images to obtain the plurality of frames of sliced ​​images includes: performing region segmentation on the plurality of frames of optimized target images based on a sliding window to obtain the plurality of frames of sliced ​​images.

[0069] Specifically, the window size and step size of the sliding window are set, and then the sliding window is set. Based on the sliding window, several frames of optimized target images are segmented to obtain several frames of sliced ​​images. The image slicing technology ensures that each slice contains enough detail information to avoid small targets being ignored in large-resolution images. At the same time, the slice super-inference method decomposes the large-resolution image into several small blocks, reducing the amount of calculation for a single inference and improving the inference efficiency.

[0070] Target detection module 24: used to construct an improved YOLOv8 model, perform target detection on a plurality of frame slice images based on the improved YOLOv8 model, and obtain corresponding target detection results;

[0071] In the specific implementation process of the present invention, the structure of the improved YOLOv8 model includes a backbone network, a neck network and a detection head.

[0072] Furthermore, the backbone network adopts a cross-stage local network, which includes several convolutional layers and a feature pyramid network, the neck network includes a feature pyramid network, a path aggregation network and an upsampling layer, and the detection head includes a small target detection layer.

[0073] Furthermore, the feature pyramid network is used to convey strong semantic features from top to bottom, the path aggregation network is used to convey positioning features from bottom to top, and the detection head is used to perform target detection prediction at different feature levels.

[0074] Furthermore, several frames of sliced ​​images generated by optimizing and slicing several frames of preprocessed target images by a data enhancement method based on hybrid enhancement are used as training data sets, and the training data sets are input into the improved YOLOv8 model for training.

[0075] Specifically, facing the complexity of the marine environment, such as variable lighting conditions, wave interference and target morphological diversity, it is often difficult to obtain satisfactory recognition results by training models based solely on raw image data. In order to improve the generalization ability and robustness of the model, data enhancement technology is widely used in the training process of deep learning models. In response to the above problems, a high-performance data enhancement method based on hybrid enhancement is designed. First, the basic data enhancement method is used to generate more samples by performing various transformations and processing on the training data. Its main purpose is to increase the diversity of the data set, prevent model overfitting, and improve the generalization ability of the model in different environments. First, the network is trained, and after the training is completed, the neural network is used to assist in segmentation. Then, the segmentation results are classified and screened, and the segmentation errors are manually corrected. Objects with obvious features and completeness are selected to form a database for subsequent data enhancement processing. Another expansion method is to directly search for images of objects such as buoys, sailboats, speedboats, etc. in the search engine, and segment and save them using the neural network according to the above method for subsequent processing. After obtaining the segmented objects, the designed data enhancement method is to use a script to paste the segmented images obtained in the previous step into the data set according to the size and category as much as possible according to the characteristics of the original data set. For example: ships in the distance are smaller and closer to the coastline; ships in the near distance are larger and located below the coastline; the coast is above the coastline, etc. The above data enhancements are all performed randomly, so it is necessary to manually judge whether the generated images are reasonable in the end, and then put them into the training set for training. After that, find other international open source data sets, filter out pictures with obstacles such as ships and buoys, and add them to the training set for training. In the training set, a data enhancement method based on hybrid enhancement is used. First, traditional image changes such as random cropping, rotation, and flipping are used to increase the diversity of training samples and enhance the robustness of the model. Then, according to the particularity of the sea surface target recognition task, a hybrid enhancement method is used in the training process to linearly superimpose two images at random proportions and adjust the labels accordingly. Generating new training samples in this way helps the model better learn the relationship between different targets and helps improve the recognition performance in complex sea environments, especially for the case of multiple target overlaps. This method can effectively solve the target recognition problem in the case of multiple target confusion. At the same time, the Cutout method and the Mosaic principle are used to randomly cover parts of the image, simulate occlusion, and enhance the recognition ability of the model when the target is partially occluded. Since the YOLO model does not match the image scale in the problem to be solved, and in order to enhance the generalization ability of the model, the Mosaic principle is used to stitch together multiple images to enhance the detection ability of targets at different scales and increase the diversity of the data set.Adversarial training is also used in the training process. By generating adversarial samples to train the model, it can resist potential adversarial attacks and enhance the robustness of the model. In the recognition of sea surface targets, images with noise or other interference factors can be generated for training to improve the performance of the model in harsh environments. Through the above data enhancement method based on hybrid enhanced intelligence, the model can use adversarial training methods in more diverse data sets to achieve better performance.

[0076] The second problem for sea surface target recognition is that the YOLO (You Only Look Once) neural network has a poor recognition effect on small targets. The YOLO neural network is an object detection system for computer vision tasks. The YOLOv8 model is an object detection model based on the YOLO neural network. There are many small targets at long distances in the data set. First, the model architecture of YOLOv8 is improved. After analyzing the small targets in the data set, it is found that these small targets are caused by the long distance of the lens. Such targets are relatively complex, usually the side or front of the ship, and there are also some long-distance buoys, so it is difficult to recognize such targets. In response to this situation, a detection layer is added on the basis of YOLOv8 to improve the feature extraction ability of small targets. Since the sample size of small targets is small and the sampling multiple of YOLOv8 is relatively large, it is difficult for the model to learn the feature information of small targets for deeper target feature maps. Based on the original network model of YOLOv8, the feature extraction network is modified to achieve small target detection by adding upsampling layers and convolution layers. In addition, parameters such as the size of the convolution kernel are modified according to the target features to better recognize the distant sailboats in the data set. The YOLOv8 model uses a cross-stage local network as the backbone network, which consists of a series of convolution layers and a feature pyramid network. The output feature map of the backbone network will be passed to the subsequent processing layer. The network structure design of the neck network applies the structure of the feature pyramid network plus the path aggregation network. The feature pyramid network layer conveys strong semantic features from top to bottom, while the path aggregation network tower conveys positioning features from bottom to top. The detection head is responsible for predicting the bounding box, category and confidence of the target. Detection heads of different scales are responsible for target detection prediction at different feature levels. Compared with the original model, the input resolution of the backbone network is increased to 1056, the neck network part adds an upsampling layer, sets a new convolution kernel size for small targets, and the detection head part is changed from three layers to four layers, expanding the small target detection layer. In order to improve the generalization ability of the model and enhance the detection performance of the small target detection layer, the modified model is pre-trained on the Common Objects in Context (COCO) dataset, and the complexity of the model is reduced by model pruning. The pre-trained model is transferred to the dataset for training to obtain a trained improved YOLOv8 model. Several frames of slice images are input into the trained improved YOLOv8 model to obtain the target detection result corresponding to each frame of slice image.

[0077] The target detection fusion module 25 is used to fuse the target detection results of several frame slice images to obtain the overall target detection result.

[0078] In the specific implementation process of the present invention, the target detection results of several frame slice images are fused to obtain the overall target detection result, including: in the process of fusing the target detection results of several frame slice images, the target detection results of all frame slice images are subjected to overlapping detection frame removal processing based on the non-maximum suppression method to obtain several target detection frames, and the overall target detection result is generated based on the several target detection frames.

[0079] Specifically, the target detection method mainly uses low-resolution images (640x480) on the COCO dataset. In these datasets, the targets are relatively large objects and have large pixel coverage, which covers 60% of the image height on average. The models pre-trained using these datasets provide very successful detection performance for similar inputs. However, the accuracy in small object detection tasks is significantly lower in high-resolution images. The image slicing assisted super reasoning method is to divide the image into several regions through a sliding window, predict each region separately, then merge the prediction results of each region with the prediction results of the entire image, and finally filter with non-maximum suppression. The pixel area of ​​small targets is relatively large relative to the image input to the network. For the analysis of the sea surface target recognition task, the image slicing assisted super reasoning method can effectively realize the recognition of small targets. In order to improve the detection performance of small targets at ultra-long distances, the image slicing assisted super reasoning method based on spatial attention is adopted. This method realizes efficient and accurate sea surface target recognition by introducing spatial attention mechanism and image slicing technology. By performing target detection on each slice image separately and introducing spatial attention mechanism, the model can better pay attention to the location of small targets and improve detection accuracy. Finally, the detection results of all slices are fused, and the non-maximum suppression method is used to remove the overlapping detection frames, that is, the slice detection results are reconstructed into the original image to form a complete target detection output. According to the confidence threshold of the detection frame, the low-confidence detection results are filtered and the high-confidence target detection frame is retained. After the introduction of the spatial attention mechanism, the model can pay more attention to the location of small targets more accurately, improving the detection accuracy of small targets. The image slicing technology ensures that each slice contains enough detail information to avoid small targets being ignored in large-resolution images. At the same time, the slice super-inference method decomposes the large-resolution image into several small blocks, reducing the computational amount of a single inference and improving the inference efficiency. The image slicing assisted super-inference method based on spatial attention effectively improves the performance of YOLOv8 in sea surface target recognition by introducing the spatial attention mechanism and image slicing technology. This method performs well in improving the detection accuracy of small targets, enhancing robustness in complex backgrounds, and improving computational efficiency. The image slicing super-inference method based on spatial attention has a very significant improvement in improving the recognition probability of small targets. However, for coastal targets, if the intersection-over-union ratio is set to a large value, non-maximum suppression sometimes does not work properly. For this reason, this method is only used when identifying obstacle labels, and the prediction results from the full image are used when identifying the coast. The fusion of the two can obtain the best result.

[0080] In the embodiment of the present invention, a data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image, and a high-performance data enhancement method is used to solve the difficulties such as multi-target overlap. An improved YOLOv8 model is constructed, and a small target detection layer is set to detect small targets on the sea surface. Target detection is performed on several frames of slice images based on the improved YOLOv8 model to obtain corresponding target detection results, and the target detection results of several frames of slice images are fused to achieve slice hyper-reasoning based on the required target location. While achieving high performance, the algorithm complexity can be effectively reduced, and the detection of offshore and coastal targets can be achieved at a lower cost.

[0081] A person skilled in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0082] In addition, the above is a detailed introduction to a sea surface target recognition method and system provided by an embodiment of the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for identifying sea surface targets, characterized in that: The method comprises: Acquire several frames of target images of the sea surface real scene video, and preprocess each frame of the target image to obtain several frames of preprocessed target images; The data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image to obtain several frames of optimized target images; Slicing a plurality of frames of optimized target images to obtain a plurality of frames of sliced ​​images; Construct an improved YOLOv8 model, perform target detection on several frames of slice images based on the improved YOLOv8 model, and obtain corresponding target detection results; The target detection results of several frames of slice images are fused to obtain the overall target detection result.

2. The method for identifying sea surface targets according to claim 1, characterized in that: The preprocessing of each frame of the target image to obtain a plurality of frames of preprocessed target images includes: Perform denoising on each frame of the target image to obtain several frames of denoised target images; Image enhancement processing is performed on each frame of the target image after denoising to obtain several frames of target images after preprocessing.

3. The method for identifying sea surface targets according to claim 1, characterized in that: The data enhancement method based on hybrid enhancement is used to optimize the image data of each frame of the preprocessed target image to obtain several frames of optimized target images, including: Perform segmentation processing on each frame of the preprocessed target image to obtain several frames of segmented images; Performing preliminary image enhancement processing on a plurality of frames of segmented images to obtain a plurality of frames of preliminary enhanced segmented images; Linearly superimposing several frames of preliminary enhanced segmented images to obtain preliminary enhanced segmented images after linear superposition of several frames; Based on the mosaic principle and Cutout method, the preliminary enhanced segmented images after linear superposition of several frames are spliced ​​and optimized to obtain several frames of optimized target images.

4. The method for identifying sea surface targets according to claim 1, characterized in that: The step of slicing a plurality of frames of optimized target images to obtain a plurality of frames of sliced ​​images includes: Based on the sliding window, several frames of optimized target images are segmented into regions to obtain several frames of sliced ​​images.

5. The method for identifying sea surface targets according to claim 1, characterized in that: The structure of the improved YOLOv8 model includes a backbone network, a neck network and a detection head.

6. The method for identifying sea surface targets according to claim 5, characterized in that: The backbone network adopts a cross-stage local network, which includes several convolutional layers and a feature pyramid network. The neck network includes a feature pyramid network, a path aggregation network and an upsampling layer. The detection head includes a small target detection layer.

7. The method for identifying sea surface targets according to claim 6, characterized in that: The feature pyramid network is used to convey strong semantic features from top to bottom, the path aggregation network is used to convey positioning features from bottom to top, and the detection head is used to perform target detection prediction at different feature levels.

8. The method for identifying sea surface targets according to claim 5, characterized in that: A plurality of frames of sliced ​​images generated by optimizing and slicing a plurality of frames of preprocessed target images by a data enhancement method based on hybrid enhancement are used as a training data set, and the training data set is input into an improved YOLOv8 model for training.

9. The method for identifying sea surface targets according to claim 1, characterized in that: The target detection results of several frames of slice images are fused to obtain the overall target detection result, including: In the process of fusing the target detection results of several frame slice images, the target detection results of all frame slice images are subjected to overlapping detection frame removal processing based on the non-maximum suppression method to obtain several target detection frames, and the overall target detection result is generated based on the several target detection frames.

10. A sea surface target recognition system, characterized in that: The system comprises: Image acquisition and preprocessing module: used to acquire several frames of target images of the sea surface real scene video, and preprocess each frame of the target image to obtain several frames of preprocessed target images; Image data optimization module: used to optimize the image data of each frame of the pre-processed target image by using a data enhancement method based on hybrid enhancement, and obtain several frames of optimized target images; Image slicing module: used to slice several frames of optimized target images to obtain several frames of sliced ​​images; Target detection module: used to construct an improved YOLOv8 model, perform target detection on several frame slice images based on the improved YOLOv8 model, and obtain corresponding target detection results; Target detection fusion module: used to fuse the target detection results of several frame slice images to obtain the overall target detection result.

Images