Sea surface small target detection method based on multi-feature image preprocessing

Abstract

The invention discloses a sea surface small target detection method based on multi-feature image preprocessing, and the method comprises the steps: extracting three features from a radar echo sequence, and generating a corresponding feature matrix; preprocessing the feature matrix to generate a high-resolution color picture; a space attention module and a channel attention module are introduced into the YOLOv8 model, the improved model is trained, and the trained model is used for sea surface small target detection; through the improved YOLOv8 model, the problems of model mismatching, low feature information utilization rate and marking difficulty in sea surface small target detection are solved, the detection performance is improved, lightweight processing is carried out, the sea condition change and feature weight processing capability is enhanced, and the controllability of the false alarm rate is realized. The method has high detection accuracy and robustness.

Description

Technical Field

[0001] The present invention belongs to the field of image detection technology, and further relates to a method for detecting small sea surface targets based on multi-feature image preprocessing in the field of remote sensing target detection technology. The present invention can be used to detect small slow-moving sea surface targets in real time and with high precision. Background Art

[0002] Detecting small slow targets on the sea surface is one of the key tasks of sea detection radar. These targets include channel buoys, small ships, etc., and their echoes on the radar are usually weak. Under the observation of the "double-high" radar system, sea clutter shows obvious non-stationary, non-uniform and non-Gaussian characteristics, and the motion model of small targets on the sea surface is complex and difficult to detect directly. Especially in the case of high resolution, low incident angle and high sea conditions, sea spikes occur frequently, making it difficult for traditional detection technology to effectively detect these low-speed floating small targets. Feature detection can better solve the difficulties encountered in energy accumulation detection and statistical model distribution detection. With the advancement of technology, multi-feature joint detectors have gradually developed and become mainstream, and their detection performance and robustness have been significantly improved. The introduction of three-feature detectors of average amplitude, relative Doppler vector entropy and relative Doppler peak height makes the advantage of multi-dimensional feature joint detection obvious. The continuous advancement of deep learning technology has made it more efficient to detect and identify target objects in unstructured data such as images and videos. At present, the optical images of sea surface targets are detected using the You Only Look Once (YOLO) algorithm. By enhancing data and adding detection layers, the detection capability of small targets is improved. However, when the optical conditions are poor, the detection effect is poor and all-weather detection cannot be achieved. In addition, optical remote sensing images are easily affected by lighting, clouds and atmospheric conditions, have poor robustness, and have low detection accuracy for single features. Although the improved YOLO algorithm improves the average accuracy of the model and reduces the amount of model calculation, it can only effectively detect ship targets in certain specific environments, has low utilization of image information, and has poor detection effect.

[0003] Beijing Institute of Technology disclosed a method for recognizing small sea targets based on visual saliency in its patent application “A method for recognizing sea targets based on visual saliency” (application number CN 202310826817.3 application publication number CN 117079097 A). The implementation steps of this method are as follows: (1) Collect and enhance sea target images and construct a data set. The data is processed by photometric and geometric distortion to increase the robustness to illumination, noise and complex background, and finally the data set is divided into a training set, a validation set and a test set; (2) A YOLO detection algorithm with fused features is constructed, and the fusion of features of different scales is enhanced by weighted parameters to improve the detection speed and accuracy; (3) The visual saliency model is used to extract the region of interest, segment the image and pre-screen the target area, and finally the trained YOLO algorithm is used for detection to ensure real-time performance and accuracy. This method comprehensively utilizes the advantages of the YOLO detection algorithm with fast detection speed and the advantages of multi-feature fusion to improve detection accuracy, thereby improving the speed and accuracy of sea target detection. However, this method still has the following shortcomings: since the echo of slow small targets is weak and the movement speed is slow, the single feature information is often masked by changes in sea conditions. The visual saliency proposed by this method can only enhance the unmasked information, but the effect on the occluded information is poor. Therefore, the detection accuracy of a single feature is low.

[0004] The Institute of Marine Instruments and Meters of Shandong Academy of Sciences disclosed a method for detecting ships on the sea surface under poor optical conditions in its patent application "A method for identifying ships at sea based on deep learning" (application number CN 202410300084.4 application publication number CN 118053102 A). The implementation steps of this method are: (1) data acquisition, extracting images frame by frame through radar video data to form a data set; (2) preprocessing the data set and improving the YOLOv8 network, including replacing part of the network structure, improving the loss function, and adding a channel attention mechanism; (3) based on the improved YOLOv8 network, constructing and training a ship recognition model, and finally realizing the recognition of ship types. Although this method improves the accuracy and stability of marine ship recognition by introducing deformable convolution and attention mechanism, and reduces the amount of calculation, it improves the accuracy to a certain extent. However, this method still has the following shortcomings: since this method uses the image features of radar time-frequency data, it cannot reflect the physical features, and the image processing of feature information will lead to information loss, thereby reducing the utilization rate of image information.

[0005] Nanjing University of Posts and Telecommunications disclosed a method for detecting small targets on the sea using air-to-ground optical images in its patent application "Small Target Detection Algorithm for UAV Images Based on Sea Surface Rescue" (application number CN 202311567359.2 application publication number CN 117576591 A). The implementation steps of this method are: (1) collecting image data of sea and beach scenes and dividing them into training sets and validation sets; (2) performing data enhancement on the sea area small target data set to improve the robustness of the model; (3) designing and using spatial attention modules and channel attention modules, training the small target data set, saving the best model, and performing detection on the test set to obtain results. Although this method performs well under good optical conditions, it still has the following shortcomings: Since this method uses air-to-ground optical images for detection, changes in optical conditions make it difficult to detect small targets on the sea surface all day long. In poor light conditions and extreme weather, the existing YOLOv8 technology is difficult to apply. Summary of the invention

[0006] The purpose of the present invention is to address the shortcomings of the above-mentioned prior art and propose a method for detecting small targets on the sea surface based on multi-feature image preprocessing, aiming to solve the problems of low detection accuracy of single features, low utilization of image information and difficulty in all-weather detection of small targets on the sea surface in existing projects.

[0007] The technical idea for achieving the purpose of the present invention is: the present invention introduces a spatial attention module and a channel attention module into the YOLOv8 model, so that the weights of the three features extracted from the radar data can change with the surrounding environment, and flexibly assign the optimal three-feature weight solution for different sea conditions. The feature information under different environments is fully utilized to solve the problem of low accuracy of traditional feature map detection. The present invention extracts three features from the radar echo data, namely, relative average amplitude RAA (Relative Average Amplitude), relative Doppler vector entropy RVE (Relative Vector Entropy) and relative Doppler peak height RDPH (Relative Doppler Peak Height), and forms matrices with these three features, which are feature data based on observation time and distance units. Next, these feature matrices are normalized and their values ​​are scaled to the red, green and blue RGB (Red Green Blue) range between 0 and 255. Then, the normalized three-dimensional features are mapped to the RGB channels respectively: RAA is mapped to the R channel, RDPH is mapped to the G channel, and RVE is mapped to the B channel. A picture is generated according to these mapping values, and the generated picture is deformed, the original 1×14 matrix is ​​changed into a 2×7 matrix, each pixel block represents the three eigenvalues ​​of a distance unit, and these pixel blocks are enlarged 100 times, and finally a 200×700 pixel picture is generated. By compressing multiple feature information into a feature image, the model is trained by using multiple feature information as joint input, and after the image generation is completed, the image is annotated, and the model training is performed on this basis. The above steps solve the problem of low utilization rate of image information in the prior art. In addition, the present invention uses the feature information extracted from the radar echo as a feature channel to replace the original optical channel, and uses the characteristic of radar to extract information from the electromagnetic echo, which is not affected by optical conditions, thereby solving the problem of excessive dependence of the prior art on the optical environment and realizing all-weather detection.

[0008] The steps of the present invention include the following:

[0009] Step 1, extract three features from the radar echo sequence and generate the corresponding feature matrix;

[0010] Step 2, normalize each feature matrix;

[0011] Step 3, mapping the normalized feature matrix to the optical channels respectively, and generating a high-resolution color image after deformation and amplification;

[0012] Step 4: label each small sea surface target in each high-resolution color image, and form a training set with all high-resolution color images and their labels;

[0013] Step 5, introduce the spatial attention module and the channel attention module into the YOLOv8 model to obtain an improved model, train the improved model to obtain a trained model; Step 6, adopt the same method as steps 1-3, obtain the radar echo data to be detected, pre-process it and generate a high-resolution color picture, input the picture into the trained model for detection, and output the result of small target detection on the sea surface.

[0014] Step 6: Use the same method as steps 1-3 to obtain the radar echo data to be detected, pre-process it and generate a high-resolution color image, input the image into the trained model for detection, and output the result of small target detection on the sea surface.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] First, the present invention introduces the spatial attention module and the channel attention module into the YOLOv8 model, so that the weights of the three features extracted from the radar data can change with the surrounding environment, and flexibly assign the optimal three feature weight solutions for different sea conditions, thereby solving the problem of low accuracy of traditional feature map detection. This allows the present invention to make full use of the feature information in different environments and improve the accuracy of small target detection on the sea surface.

[0017] Second, the present invention compresses three types of feature information into a feature image and uses multi-feature information as joint input to train the model, thereby overcoming the shortcoming of the prior art in low utilization of image information. The present invention improves information utilization and reduces model training costs, thereby effectively improving the practicability of small surface target detection.

[0018] Third, the present invention uses the characteristic information extracted from the radar echo as the characteristic channel to replace the original optical channel, and utilizes the characteristic of radar extracting information from electromagnetic echoes, which is not affected by optical conditions, thus overcoming the problem of the prior art being too dependent on the optical environment. This enables the present invention to detect small targets on the sea surface in real time and in all-weather conditions at night when there is insufficient light and in bad weather, greatly improving the application value of this method in engineering practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the present invention;

[0020] Figure 2 It is a data preprocessing flow chart of the present invention;

[0021] Figure 3 It is a schematic diagram of the YOLOv8-CHINS structure of the present invention;

[0022] Figure 4 Schematic diagram of the channel attention model structure of the present invention;

[0023] Figure 5 Schematic diagram of the spatial attention model structure of the present invention;

[0024] Figure 6 This is a comparison chart of the detection probabilities of the detector proposed in the simulation experiment of the present invention and the existing detector. DETAILED DESCRIPTION

[0025] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0026] Reference Figure 1 , further describing the implementation steps of the embodiment of the present invention.

[0027] Step 1: Extract three features from the radar echo sequence and generate the corresponding feature matrix.

[0028] The embodiment of the present invention extracts three features RAA (relative average amplitude), RDPH (relative Doppler peak height) and RVE (relative vector entropy) from the radar echo sequence s(n), and respectively composes them into feature matrices S0, S1 and S2, and these three matrices are all m×n matrices, where m=1024 / N is the number of features, N is the observation time step, and n=14 is the number of distance units.

[0029] The three features of RAA, RDPH and RVE are obtained by the following steps.

[0030] Step 2.1, use the following formula to calculate the average amplitude of the detection unit and the reference unit in the radar echo sequence and then extract the relative average amplitude feature of the target unit:

[0031]

[0032] Where RAA(z) represents the relative average amplitude feature, ∑· represents the summation operation, n=1,2,…,N, N represents the length of the radar echo sequence, |·| represents the absolute value operation, z(n) represents the target unit, and Θ(·) represents the function used to extract ordered statistics, such as the mean, median or other related statistics. k (n) represents a reference unit, k=1, 2, ..., K, and K represents the total number of reference units.

[0033] Step 2.2, calculate the Doppler amplitude value of the radar echo sequence according to the following formula:

[0034]

[0035]

[0036] Among them, Z(f d ) represents the dth Doppler frequency f in the radar echo sequenced The Doppler amplitude value of the radar echo sequence is represented by exp, which represents the exponential operation with the natural constant e as the base, π represents the circumference of the circle, and PRP represents the pulse repetition period of the radar echo sequence.

[0037] Step 2.3, calculate the Doppler peak height and Doppler shift according to the following formula:

[0038]

[0039] Where Peak(z) represents the Doppler peak height, Indicates the Doppler shift.

[0040] Step 2.4: Extract the relative Doppler peak height feature according to the following formula:

[0041] D=[-δ 1 ,-δ 2 ]∪[δ 2 ,δ 1 ]

[0042]

[0043] Where D represents the set of Doppler reference units, δ 1 represents the reference Doppler unit, δ 2 represents the maximum Doppler bandwidth of the target, ∪ represents the union symbol of the set, #D represents the total number of elements in the set D, and RDPH(z) represents the relative Doppler peak height feature.

[0044] Step 2.5, calculate the vector entropy of the echo sequence in the Doppler domain according to the following formula to extract the relative vector entropy feature:

[0045]

[0046]

[0047] Where VE(z) represents the vector entropy of the radar echo sequence in the Doppler domain, RVE represents the relative vector entropy feature, and log represents the logarithmic operation with base 10.

[0048] In information theory, information entropy is a key indicator used to measure the amount of information. Based on this concept, the VE vector entropy of the Doppler spectrum is introduced, which helps to distinguish the echo sequence performance of pure sea clutter and the target range unit in the Doppler domain. In general, the echo sequence of pure sea clutter usually has a higher vector entropy because the Doppler frequency band is wide and scattered. Relatively speaking, in the echo sequence containing the target, the Doppler frequency band is narrow and concentrated, so its vector entropy is often lower. By comparing the Doppler domain vector entropy of the echo sequence of the unit to be tested with the Doppler domain vector entropy of the reference unit echo sequence, the RVE relative vector entropy can be calculated. This ratio helps to achieve the characteristic of constant false alarm rate.

[0049] Step 2, refer to Figure 2 In the normalization processing part, each feature matrix is ​​normalized, that is, the three eigenvalues ​​are scaled to the following range according to the following formula:

[0050]

[0051] Among them, S represents the feature matrix, S max and S min Represent the maximum and minimum values ​​in the S matrix respectively.

[0052] Step 3: Map the normalized feature matrix to the optical channels respectively, and generate a high-resolution color image after deformation and amplification.

[0053] Reference Figure 2 Data preprocessing flowchart, detailing the steps for generating high-resolution color images.

[0054] In the first step, the normalized feature matrix of 1×14 dimensions is transformed into a matrix of 2×7 dimensions.

[0055] In the second step, the deformed matrix is ​​mapped to the red, green, and blue RGB channels, the feature RAA is mapped to the R channel, the feature RDPH is mapped to the G channel, and the feature RVE is mapped to the B channel; each group of three feature matrices generates a 2×7 pixel block.

[0056] The third step is to enlarge each pixel block by 100 times to generate a 200×700 pixel high-resolution color image corresponding to each observation time.

[0057] Step 4: label each small sea surface target in each high-resolution color image, and form a training set with all high-resolution color images and their annotations.

[0058] Step 5: Introduce the spatial attention module and the channel attention module into the YOLOv8 model to obtain the improved YOLOv8-CHINS model, and train the improved model to obtain the trained model.

[0059] Step 5.1, introduce the spatial attention module and channel attention module into the YOLOv8 model to obtain the improved YOLOv8-CHINS model.

[0060] The structure and parameters of the YOLOv8 model are as follows: the 0th and 1st layers are convolutional layers, with output channels of 64 and 128 respectively, the convolution kernel size of 3, and the step size of 2; the 2nd layer is a C2f module, with an output channel of 128, connected by shortcut; the 3rd layer is a convolutional layer, with an output channel of 256, the convolution kernel size of 3, and the step size of 2; the 4th layer is a C2f module, with an output channel of 256, connected by shortcut; the 5th layer is a convolutional layer, with an output channel of 512, the convolution kernel size of 3, and the step size of 2; the 6th layer is a SE_C2f module, with an output channel of 512, connected by shortcut; the 7th layer is a convolutional layer, with an output channel of 1024, the convolution kernel size of 3, and the step size of 2; the 8th layer is a SE_C2f module, with an output channel of 1024, connected by shortcut; the 9th layer is a SPPF layer, with an output channel of The number of channels is 1024, and the pooling kernel size is 5; the 10th layer is the CBAM layer, the output channel is 1024, and the convolution kernel size is 7; the 11th layer is the upsampling layer, using the nearest neighbor interpolation, and the scale_factor is set to 2; the 12th layer is the Concat layer, which is used to concatenate the output of the 6th layer with the output of the 11th layer in this layer; the 13th layer is the C2f module, the output channel is 512, and there is no shortcut connection; the 14th layer is the convolution layer, the output channel is 1024, the convolution kernel size is 3, and the step size is 2; the 15th layer is the convolution layer, the output channel is 512, the convolution kernel size is 3, and the step size is 2; the 16th layer is the Concat layer, which is used to concatenate the output of the 8th layer with the output of the 15th layer in this layer; the 17th layer is the C2f module, the output channel is 512, and there is no shortcut connection; the 18th layer is the Detect layer, which is used to complete the final detection task.

[0061] The present invention improves the input image so that it can accommodate more feature information. At the same time, the spatial attention module and channel attention module that are more suitable for the present invention are introduced into the original YOLOv8 network architecture.

[0062] Reference Figure 3 The YOLOv8-CHINS structure diagram of the present invention further describes the introduction of the spatial attention module and the channel attention module in the YOLOv8 model.

[0063] The channel attention module is set between the 10th and 11th layers of the YOLOv8 model; it focuses on the channel weight information of the feature map, extracts the global information of each channel through global average pooling and global maximum pooling, and then calculates the attention weight of the channel through a fully connected layer or a convolutional layer; the weight map is used to weight the channel dimension of the feature map, thereby enhancing the model's response to key channels.

[0064] The spatial attention module is set between the 6th and 7th layers of the YOLOv8 model; it focuses on the spatial information in the feature map, and generates spatial attention weights by performing average pooling and maximum pooling on the channel dimensions of the input feature map and performing convolution operations on the pooled features; the weight map is used to weight each position of the input feature map, so that the model pays more attention to important spatial areas.

[0065] The attention module of the convolutional neural network adopted by the attention module of the embodiment of the present invention is a structural unit designed to improve the performance of the convolutional neural network. This goal is achieved by fusing the two mechanisms of channel attention and spatial attention, thereby enhancing the network's perception of information without increasing the complexity of the network structure.

[0066] Reference Figure 4 , further describe the channel attention module.

[0067] The channel attention module is composed of the following parts in series: the data with input matrix length C is sent to the first part as CBS module, and the output size is 128 channels; the second part is C2f module, the output size is 256 channels, the convolution kernel size is 3, and the step size is 3; the third part is the maximum pooling module, which includes global maximum pooling and global average pooling, the step size is 2, the output size is 256 channels, and the output data is output to the next part through the multi-layer perceptron; the fourth part is the feature pyramid module and Sigmoid activation function, the output is 512 channels, the output size is 256 channels, and the input part and the output channel attention data M are combined. C Splicing, the final output size is 128 channels of weight coefficient F C .

[0068] The channel attention mechanism corresponds to the weight calculation of RAA, RDPH and RVE, and the spatial attention mechanism is the weight calculation of sea condition changes. In this way, through a large amount of data training, the weights of the three features extracted from the radar data can change with the surrounding environment. Instead of seeking a balanced solution under different sea conditions, the optimal three-feature weight solution is intelligently assigned for different sea conditions. It can be applied to a variety of CNN architectures to enhance the model's ability to process complex data. The channel attention module evaluates the importance of each channel to determine their contribution to the final feature. This module first performs global average pooling on the feature map to extract the global feature description of each channel. These descriptions are converted into channel weights through a two-layer fully connected network, and then the original feature map is adjusted by these weights to refine the feature expression and improve the feature discrimination.

[0069]

[0070] The symbol σ is used to represent the Sigmoid function, W M ∈R c×C / r ,W 1 ∈R C×C / r , and these two matrices are shared in the multi-layer perception P. 0 Before the calculation of , the ReLU activation function is introduced to increase nonlinearity.

[0071] Reference Figure 5 , further describe the spatial attention module.

[0072] The spatial attention module is composed of the following parts in series: the first part is a CA module, which accepts data with a matrix length of W×C, and the output size is 128 channels; the second part is a conv module, which accepts data with a size of W×H, and maps it to a W×H matrix after passing through a 1×1 convolution kernel, with an output size of 256 channels, a convolution kernel size of 3, and a step size of 2; the third part is a maximum pooling module, which first concatenates the first part input and the second part output through concat as input, and outputs 512 channels; the fourth part is an SPPF module and an activation function, which outputs the spatial attention data M after passing through a 1×1 convolution layer. S Added to the input features, the output size is 128 channels of weight coefficient F S .

[0073] The spatial attention module focuses on evaluating the spatial importance of each pixel in the image. This module receives the output of the channel attention module and calculates the average and maximum values ​​of each channel by performing max pooling and average pooling operations. These calculated values ​​are concatenated and fed into the fully connected layer to generate a weight vector representing the spatial importance of each pixel. This spatial weight vector is multiplied by the adjusted feature map to generate the final feature map adjusted by spatial attention. This process helps to distinguish the spatial importance of different pixels more accurately, further improving the expressiveness of the features.

[0074]

[0075] Where σ represents the Sigmoid function, f 7×7 It is a convolution kernel of size 7×7.

[0076] The spatial attention module can adaptively adjust the weights of different feature channels and pixel spatial information in the feature map, automatically select the best feature combination for different tasks and datasets, and thus significantly improve the overall performance of the model.

[0077]

[0078]

[0079] Among them, F represents the processed intermediate feature map, F∈R c×H×W Through the channel attention module, a one-dimensional channel attention map M is generated O , M O ∈R c×1×1 At the same time, the spatial attention module generates a two-dimensional spatial attention map M s , M s ∈R 1×H×W This configuration ensures the flexibility and accuracy of the model in dealing with complex sea conditions.

[0080] Step 5.2: Train the improved model.

[0081] The training set is input into the improved model, the gradient of the loss function is calculated through the back propagation algorithm, and the parameters of the model are iteratively updated until the loss function of the model converges to obtain a trained model.

[0082] The loss function of the model is as follows:

[0083]

[0084] B gt =(x 0 gt ,y 0 gt ,w0 gt ,h 0 gt )

[0085] Among them, CIoU-Loss represents the loss function of the model, x 0 gt ,y 0 gt Respectively represent the coordinates of the center of the annotation box of each small sea target in the training set, w 0 gt ,h 0 gt They represent the width and height of the prediction box output during the improved model training, w gt ,h gt Respectively represent the width and height of the target annotation box in the training set, b gt represents the center point of the annotation box, b represents the center point of the prediction box output during the improved model training, and B gt represents the loss function matrix of the prediction box output during the training of the improved model, B represents the loss function matrix of the labeled box of each small target on the sea surface in the training set, ρ(·) represents the Euclidean distance, c represents the diagonal length of the minimum circumscribed rectangle covering the two boxes, v represents the parameter used to evaluate the consistency of the aspect ratio between the prediction box and the actual box, and α represents the adjustment factor used to adjust the weight based on the IoU value. Its function is to give a larger weight when the IoU value between the prediction box and the target box is higher.

[0086] Step 6: Use the same method as steps 1-3 to obtain the radar echo data to be detected, pre-process it and generate a high-resolution color image, input the image into the trained model for detection, and output the result of small target detection on the sea surface.

[0087] The effects of the present invention can be specifically illustrated by the following simulation experiments.

[0088] 1. Simulation experiment conditions.

[0089] The YOLOv8 model used in the simulation experiment of the present invention is configured under the Windows 10 operating system, the CPU model is Intel(R) Core(TM) i9-14900KF3.20 GHz, the running memory is 128GB, the GPU model is NVIDIAGeForce RTX 4090, the deep learning framework is Pytorch2.2.1+cu121, the computing architecture is Cuda12.4, and the development environment is Python3.9.

[0090] The data used in the simulation experiment of the present invention is the 1993 IPIX radar data set publicly available on the Internet.

[0091] 2. Simulation content and result analysis.

[0092] The simulation experiment of the present invention uses the YOLOv8 model and the improved YOLOv8-CHINS model of the present invention to detect small targets on the sea surface on the IPIX radar data set. The results are as follows: Figure 6 shown.

[0093] Figure 6 (a) Figure 6 (b) Figure 6 (c) Figure 6 (d) The results of observing ten sets of data under the radar HH, HV, VH, and VV polarization conditions, with an observation time of 0.512s. The average detection probability of each group of multiple detectors is compared under each polarization condition.

[0094] Figure 6 The horizontal axis is the data number of each group, and the vertical axis is the average detection probability. Figure 6 The grey dotted line in the figure represents the average detection probability based on the fractal detector, the blue dotted line represents the average detection probability based on the three-feature detector, the red dotted line represents the average detection probability based on the three-feature time-frequency detector, and the green dotted line represents the average detection probability of the detector proposed in the present invention.

[0095] from Figure 6 It can be seen that in the simulation experiment of the present invention, within the observation time of 0.512 seconds, the average detection probability of the fractal-based detector under four polarization modes (HH, HV, VH, VV) is 0.329, among which the detection probability under HH polarization is the lowest, only 0.223, and the detection probability under VV polarization is the highest, which is 0.443. The average detection probability of the detector based on two-dimensional features is 0.665, which is significantly higher than that of the fractal-based detector, and the detection probabilities under HV and VH polarizations reach 0.577 and 0.736 respectively. The detector based on the three characteristics of time and frequency performs better, with an average detection probability of 0.780, especially under VV polarization, the detection probability is the highest, reaching 0.842. The detection probability of the detector proposed in the present invention in all polarization modes is higher than that of the other three detectors, and the average detection probability reaches 0.859, especially in the VH polarization mode, the detection probability is the highest, reaching 0.998, showing excellent detection performance. Obviously, the detector proposed in the present invention has demonstrated excellent detection effects in 40 sets of complete data tests, proving its superior performance and high stability.

Claims

1. A method for detecting small targets on the sea surface based on multi-feature image preprocessing, characterized in that: The three feature information extracted from the radar echo sequence is preprocessed to generate a high-resolution color image. The spatial attention module and the channel attention module are introduced into the YOLOv8 model, and the improved model is used for target detection. The steps of the detection method include the following: Step 1, extract three features from the radar echo sequence and generate the corresponding feature matrix; Step 2, normalize each feature matrix; Step 3, mapping the normalized feature matrix to the optical channels respectively, and generating a high-resolution color image after deformation and amplification; Step 4: label each small sea surface target in each high-resolution color image, and form a training set with all high-resolution color images and their labels; Step 5, introducing the spatial attention module and the channel attention module into the YOLOv8 model to obtain an improved model, and training the improved model to obtain a trained model; Step 6: Use the same method as steps 1-3 to obtain the radar echo data to be detected, pre-process it and generate a high-resolution color image, input the image into the trained model for detection, and output the result of small target detection on the sea surface.

2. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 1 is characterized in that: The steps for extracting three features from the radar echo sequence described in step 1 are as follows: In the first step, the average amplitude of the detection unit and the reference unit in the radar echo sequence is calculated using the following formula, and then the relative average amplitude feature of the target unit is extracted: Where RAA(z) represents the relative average amplitude feature, ∑· represents the summation operation, n=1,2,…,N, N represents the length of the radar echo sequence, |·| represents the absolute value operation, z(n) represents the target unit, Θ(·) represents the operator for extracting ordered statistics, and z k (n) represents the reference unit, K represents the total number of reference units; The second step is to calculate the Doppler amplitude value of the radar echo sequence according to the following formula: Among them, Z(f d ) represents the dth Doppler frequency f in the radar echo sequence d The Doppler amplitude value, exp represents the exponential operation with the natural constant e as the base, π represents pi, and PRP represents the pulse repetition period of the radar echo sequence; The third step is to calculate the Doppler peak height and Doppler shift according to the following formula: Where Peak(z) represents the Doppler peak height, represents the Doppler shift; The fourth step is to extract the relative Doppler peak height feature according to the following formula: D=[-δ1,-δ2]∪[δ2,δ1] Where D represents a set of Doppler reference units, δ1 represents the reference Doppler unit, δ2 represents the maximum Doppler bandwidth of the target, ∪ represents the union symbol of the set, #D represents the total number of elements in the set D, and RDPH(z) represents the relative Doppler peak height feature; The fifth step is to calculate the vector entropy of the echo sequence in the Doppler domain according to the following formula to extract the relative vector entropy feature: Where VE(z) represents the vector entropy of the radar echo sequence in the Doppler domain, RVE represents the relative vector entropy feature, and log represents the logarithmic operation with base 10.

3. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 1, characterized in that: The feature matrix normalization process corresponding to each feature in step 2 refers to scaling the three eigenvalues ​​to the following ranges according to the following formula: Among them, S represents the feature matrix, S 234 and S 256 Represent the maximum and minimum values ​​in the S matrix respectively.

4. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 2, characterized in that: The steps to generate a high-resolution color image described in step 3 are as follows: The first step is to transform the 1×14-dimensional normalized feature matrix into a 2×7-dimensional matrix; In the second step, the deformed matrix is ​​mapped to the red, green, and blue RGB channels, the feature RAA is mapped to the R channel, the feature RDPH is mapped to the G channel, and the feature RVE is mapped to the B channel; each group of three feature matrices generates a 2×7 pixel block; The third step is to enlarge each pixel block by 100 times to generate a 200×700 pixel high-resolution color image corresponding to each observation time.

5. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 1, characterized in that: The structure and parameters of the YOLOv8 model described in step 5 are as follows: the 0th and 1st layers are convolutional layers, with output channels of 64 and 128 respectively, the convolution kernel size is 3, and the step size is 2; the 2nd layer is a C2f module, with an output channel of 128, connected by shortcut; the 3rd layer is a convolutional layer, with an output channel of 256, the convolution kernel size is 3, and the step size is 2; the 4th layer is a C2f module, with an output channel of 256, connected by shortcut; the 5th layer is a convolutional layer, with an output channel of 512, the convolution kernel size is 3, and the step size is 2; the 6th layer is a SE_C2f module, with an output channel of 512, connected by shortcut; the 7th layer is a convolutional layer, with an output channel of 1024, the convolution kernel size is 3, and the step size is 2; the 8th layer is a SE_C2f module, with an output channel of 1024, connected by shortcut; the 9th layer is an SPPF layer, with an output The channel is 1024, and the pooling kernel size is 5; the 10th layer is the CBAM layer, the output channel is 1024, and the convolution kernel size is 7; the 11th layer is the upsampling layer, using the nearest neighbor interpolation, and the scale_factor is set to 2; the 12th layer is the Concat layer, which is used to splice the output of the 6th layer with the output of the 11th layer in this layer; the 13th layer is the C2f module, with an output channel of 512 and no shortcut connection; the 14th layer is the convolution layer, with an output channel of 1024, a convolution kernel size of 3, and a step size of 2; the 15th layer is the convolution layer, with an output of 512 channels, a convolution kernel size of 3, and a step size of 2; the 16th layer is the Concat layer, which is used to splice the output of the 8th layer with the output of the 15th layer in this layer; the 17th layer is the C2f module, with an output channel of 512 and no shortcut connection; the 18th layer is the Detect layer, which is used to complete the final detection task.

6. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 5, characterized in that: The introduction of the spatial attention module and the channel attention module into the YOLOv8 model as described in claim 5 refers to: The spatial attention module is set between the 6th and 7th layers of the YOLOv8 model; it focuses on the spatial information in the feature map, and generates spatial attention weights by performing average pooling and maximum pooling on the channel dimensions of the input feature map and performing convolution operations on the pooled features; the weight map is used to weight each position of the input feature map, so that the model pays more attention to important spatial areas; The channel attention module is set between the 10th and 11th layers of the YOLOv8 model; it focuses on the channel weight information of the feature map, extracts the global information of each channel through global average pooling and global maximum pooling, and then calculates the attention weight of the channel through a fully connected layer or a convolutional layer; the weight map is used to weight the channel dimension of the feature map, thereby enhancing the model's response to key channels.

7. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 1, characterized in that: The spatial attention module described in step 5 is composed of the following parts in series: the first part is a CA module with an output size of 128 channels; the second part is a conv module with an output size of 256 channels, a convolution kernel size of 3, and a step size of 2; the third part is a maximum pooling module, which first concatenates the input of the first part and the output of the second part through concat as input and outputs 512 channels; the third part is an SPPF module and an activation function, which outputs a weight coefficient of 128 channels after passing through a 1x1 convolution layer.

8. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 1, characterized in that: The channel attention module described in step 5 is composed of the following parts in series: the first part is the CBS module with an output size of 128 channels; the second part is the C2f module with an output size of 256 channels, a convolution kernel size of 3, and a step size of 3; the third part is the maximum pooling module with a step size of 2 and an output size of 256 channels; the fourth part is the feature pyramid module and the Sigmoid activation function, with an output size of 512 channels, an output size of 256 channels, and finally an output size of 128 channels. The weight coefficient.

9. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 1, characterized in that: The training of the improved model described in step 5 refers to inputting the training set into the improved model, calculating the gradient of the loss function through the back propagation algorithm, iteratively updating the parameters of the model until the loss function of the model converges, and obtaining a trained model.

10. The method for detecting small sea targets based on multi-feature image preprocessing according to claim 9, characterized in that: The loss function of the model is as follows: Among them, CIoU-Loss represents the loss function of the model, Respectively represent the coordinates of the center of the annotation box of each small sea target in the training set, They represent the width and height of the prediction box output during the improved model training, w gt ,h gt Respectively represent the width and height of the target annotation box in the training set, b 89 represents the center point of the annotation box, b represents the center point of the prediction box output during the improved model training, and B gt represents the loss function matrix of the prediction box output during the training of the improved model, B represents the loss function matrix of the labeled box of each small target on the sea surface in the training set, ρ(·) represents the Euclidean distance, c represents the diagonal length of the minimum circumscribed rectangle covering the two boxes, v represents the parameter used to evaluate the consistency of the aspect ratio between the prediction box and the actual box, and α represents the adjustment factor used to adjust the weight based on the IoU value. Its function is to give a larger weight when the IoU value between the prediction box and the target box is higher.

Images