A method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images

Abstract

The present invention relates to the technical field of remote sensing image target detection, and in particular to a method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images. The method first performs image super-resolution reconstruction and data labeling, etc., obtains a multi-scale feature map using RegNet, fuses the multi-scale feature map using a feature pyramid, and obtains multi-scale feature information. Each pixel point on the feature map is sampled, and center branch prediction and regression classification are performed on each pixel point. Non-maximum suppression is performed on the prediction results to eliminate duplicate frames. The prediction error is calculated using a loss function, and a back-propagation algorithm is used to update model parameters to obtain a ship target detection model. The method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images adds a deconvolution layer on the basis of feature extraction, effectively solves the problem of insufficient target feature information, improves the small target detection accuracy, and solves the detection model redundancy problem by pixel-by-pixel prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image target detection, in particular to a method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images. Background Art

[0002] With the continuous development of natural image target detection algorithms and the rapid advancement of remote sensing imaging technology, automated air-to-ground target detection and analysis using remote sensing images has been widely applied in maritime surveillance. Although many remote sensing image researchers have incorporated techniques from natural images into remote sensing image target detection, the high-altitude and random distribution of remote sensing images makes it difficult to accurately locate targets using the same horizontal detection frame used in natural images. Therefore, in recent years, the use of rotated detection frames has begun to be used in remote sensing image target detection. Furthermore, ship detection technology based on remote sensing images has become an important tool for maritime management.

[0003] Because ship targets in optical remote sensing images have arbitrary orientations and multi-scale characteristics, rotating box detection methods can better fit the targets and avoid background interference. Ship detection networks based on pixel-by-pixel processing of remote sensing images utilize fully convolutional networks and specific anchor boxes to directly predict the position, aspect ratio, and angle of the rotating box. This allows for target detection at various scales and can handle objects of varying sizes. However, their performance is limited when dealing with small targets. Optical remote sensing images contain a wealth of information and complex backgrounds, often resulting in small ship targets with sizes less than 5% of the input image. This small size can lead to detection issues such as insufficient image feature information and small targets being classified as background. Traditional anchor box-based methods predefine rectangular boxes of fixed size and aspect ratio within the network. This generates a large number of candidate boxes, resulting in computationally intensive training and inference, requiring more memory and computing resources. This makes the model bulky, increasing the difficulty of model training and deployment, and increasing the computational resources required for detection. This leads to redundant detection models and hinders algorithm portability and mobile platform embedding.

[0004] In summary, how to accurately detect small targets in remote sensing image target detection, and how to solve the redundancy of detection models and achieve network lightweighting are still problems to be solved in the current field of remote sensing image target detection. Summary of the Invention

[0005] Therefore, the purpose of the present invention is to provide a method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images to solve the problems of difficulty in detecting small targets and redundancy of detection models.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images includes the following steps:

[0008] Step a: Using an image slice enhancement algorithm based on a generative adversarial network to perform super-resolution reconstruction on the acquired image;

[0009] Step b: Adjust the network initial width and width multiplier coefficient to allow the RegNet network to extract input image features, obtain features at different levels, and add deconvolution operations to make the model detect small objects more accurately;

[0010] Step c: For the features at different levels, the bidirectional feature pyramid module is used to perform feature fusion using top-down and bottom-up paths, and the attention mechanism of the context refinement module is used to enhance the model's attention to the target in the region;

[0011] Step d: Sample and predict each pixel on the feature map. Use the center branch to determine the position information of the object's bounding box from the coordinates of the key points in the candidate box. Use the regression branch to estimate the size and position of the target box, generate the bounding box of the object, and use a multi-classifier to obtain the probability scores of different target categories in the area. Set a threshold. Target categories with scores below the threshold are considered background, and target categories with scores above the threshold are retained as prediction results.

[0012] Step e: Correct the bounding box using bounding box offset and scaling, then use non-maximum suppression to deduplicate the bounding box. Sorting the bounding boxes from high to low confidence, selecting the bounding box with the highest confidence one by one, and performing an intersection-union calculation on the bounding box with the highest confidence and all the remaining bounding boxes to obtain the accurate object bounding box.

[0013] Step f, calculating the error through the loss function and using the back propagation algorithm to update the model parameters to obtain a small target detection model for ships in any direction;

[0014] Step g: Use the arbitrary direction ship small target detection model to detect the input remote sensing image.

[0015] The above-mentioned method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images is for processing low-resolution remote sensing images containing ship targets. Step a includes using an image slicing enhancement algorithm based on a generative adversarial network to perform single-frame and multi-frame high-definition reconstruction of blurred targets in the input image and restore the target texture details in the image.

[0016] In the above-mentioned method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images, in order to accurately detect small targets, step b includes:

[0017] The RegNet model uses the NRB structure to extract features using multi-scale information in each block;

[0018] A deconvolution layer is added to the feature extraction network to upsample the low-resolution feature map to a high resolution through deconvolution operation.

[0019] The above-mentioned method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images, in order to reduce information loss and ambiguity, in the process of feature fusion of the bidirectional feature pyramid, in each level of the pyramid, the forward path reduces the feature map of the previous level through convolution and pooling operations, and performs feature fusion with the feature map of the next level; the reverse path enlarges the feature map of the next level through upsampling and convolution operations, and performs feature fusion with the feature map of the previous level.

[0020] In the above-mentioned method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images, step d comprises:

[0021] Remap each point in the feature map back to the original image. If the point mapped back to the original image is within the bounding box and the classification category is as expected, then the point is assigned a positive weight, otherwise it is considered negative.

[0022] The prediction of each pixel is completed by three branches. For each pixel, a square receptive field area is generated with it as the center on the feature map. Each area has a fixed scale and corresponding weight.

[0023] The classification branch outputs the probability of which category the pixel belongs to; the center point branch predicts whether the pixel is the center point of the object; finally, the bounding box regression branch predicts the bounding box of the object corresponding to the pixel; these three branches jointly process each pixel in each level and ultimately obtain all possible object boxes.

[0024] The original bounding box is corrected by the bounding box offset and scaling predicted by the network to obtain a more accurate object bounding box, and then the non-maximum suppression method is used to remove duplicate detection boxes.

[0025] The above-mentioned method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images combines classification loss, regression loss, and center loss to prevent model overfitting. The loss function of the entire network uses the cross-entropy loss function with L2 regularization to calculate the error, and uses the back-propagation algorithm to update the model parameters to obtain a small ship target detection model in any direction. The cross-entropy loss function with L2 regularization is expressed as:

[0026]

[0027] Among them, L2 represents the regularized cross entropy loss function, y i Indicates the true label of the i-th sample, true is 0, false is 1, γ i Represents the corresponding predicted probability value, N represents the number of samples, / / w / / 2 represents the sum of the squares of the weight vector, λ represents the regularization coefficient, and log represents the natural logarithm.

[0028] Beneficial effects:

[0029] First, the present invention adds a deconvolution module to the feature extraction network of the detection model. The deconvolution module increases the resolution of the feature map by upsampling the feature map, and provides richer feature information for subsequent convolution operations. Then, the convolution operation is performed to generate an output of the same size as the original feature map, so that the model can more easily capture the detailed information of small targets, thereby improving the model's detection accuracy for small targets; and in the data processing part, the image slicing enhancement algorithm based on the generative adversarial network is used to effectively solve the problem of insufficient target feature information, and realize the detection of ship targets and small targets in low-resolution remote sensing images.

[0030] Second, the present invention uses a pixel-by-pixel prediction method without predefined anchor frames, which reduces computational complexity and can perform target detection at different scales while processing objects of different scales. This greatly reduces network model parameters, achieves lightweight network models, and solves the problem of redundancy in detection models. While ensuring high efficiency, it can reduce resource consumption, improve model practicality, promote algorithm migration and integration with mobile embedded platforms, and is suitable for detection tasks on mobile devices such as mobile phones, drones, and smart cameras. The lightweight network model reduces the computational complexity of the model and improves the calculation speed and response speed of the model. The hardware resources and computing power of the mobile platform are limited, but need to meet the needs of real-time target detection. The lightweight target detection network model of the present invention can make up for this deficiency and achieve fast and accurate small target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of the method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images according to the present invention.

[0032] Figure 2 Schematic diagram for image super-resolution reconstruction comparison.

[0033] Figure 3 A comparison chart of horizontal sampling and rotation sampling.

[0034] Figure 4 This is the overall network structure diagram of the ship target detection network in any direction.

[0035] Figure 5This is the effect diagram of optical remote sensing ship target detection in any direction. DETAILED DESCRIPTION

[0036] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0037] The flowchart of the method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images in this specific embodiment is as follows: Figure 1 As shown, it includes the following steps:

[0038] Step a: Using an image slice enhancement algorithm based on a generative adversarial network to perform super-resolution reconstruction on the acquired image;

[0039] First, the input image is super-reconstructed using an image slicing enhancement algorithm based on a generative adversarial network. To increase the sample size, the original image is divided into small tiles of equal size through image slicing. These tiles are then fed into the generator network, which continuously generates high-resolution images. In this case, random noise is fed into the generator to produce a high-resolution version of the initial image.

[0040] The discriminator network, composed of a deep convolutional neural network, accepts the high-resolution images generated by the generator and the corresponding real low-resolution images, and evaluates the difference between the images generated by the generator and the real images;

[0041] The objectives of the generator and discriminator are optimized through adversarial learning. The loss functions of the generator and discriminator are expressed as follows:

[0042]

[0043] Among them, V(D,G) is the loss function between the generator G and the discriminator D, min G and max D represents minimizing the loss function of the generator and maximizing the loss function of the discriminator, It means that for the real data sample x, p data The probability distribution in (x) is weighted, and the expected value of log D(x) is taken, where p data (x) is the probability distribution of the real data, D(x) represents the discriminator’s prediction result for the data sample x, It represents the expected value of log(1-D(G(z))) for the random noise vector z sampled from the prior noise distribution p(z), with p(z) as the weight. G(z) represents the fake sample generated by the generator given the random noise vector z, such as Figure 2 Shown are comparison images before and after super-resolution reconstruction;

[0044] The datasets used in this example are DIOR, NWPUVHR-10, HRSC2016, and DOTA datasets. Except for the HRSC2016 dataset which only contains one type of remote sensing target, the other three datasets contain multiple types of remote sensing targets, including ship targets. The roLabelImg tool is used for data annotation, and an XML file is generated to record the annotation results, as shown in the following example: Figure 4 The figure shows a comparison between horizontal sampling and rotation sampling. In this embodiment, the training set and the test set are divided into 7:1 ratios.

[0045] Step b: Adjust the network initial width and width multiplier coefficient to allow the RegNet network to extract input image features, obtain features at different levels, and add deconvolution operations to make the model detect small objects more accurately;

[0046] The input image is passed through a convolutional layer for feature extraction and downsampling. The feature map is then sent to multiple basic blocks, each of which contains a series of residual connections and regularization designs to gradually improve the feature expression capability. Finally, all feature maps processed by the basic blocks are sent to a global pooling layer and connected to the final output layer. RegNet has designed a set of flexible hyperparameters, including depth, initial width, steepness, width parameter, bottleneck ratio, number of groups, etc. Through specific hyperparameter combinations, a feature extraction network that meets the requirements can be obtained. The network structure of RegNet mainly consists of three parts: stem, body and head.

[0047] stem: Contains 32 3x3 convolution kernels with a stride of 2, and also includes BN and ReLU activation functions;

[0048] Body: It consists of four stacked stages, which perform feature extraction and dimensionality reduction.

[0049] Head: All feature maps processed by the stage are average pooled, and finally a fully connected output layer is used to generate the final prediction result;

[0050] RegNet first processes the input through the stem part, then performs feature extraction and dimensionality reduction through multiple stacked stages, and finally outputs the prediction results through the head part.

[0051] In order to improve the small target detection capability, the present invention introduces a deconvolution operation based on the RegNet network. The feature map obtained by the convolution operation is remapped back to the space of the original input image through deconvolution, and the feature map is amplified to obtain the output image. The deconvolution of the present invention is implemented through two operations: interpolation and convolution. The interpolation operation is to upsample the low-resolution feature map to the same size as the input image. Bicubic interpolation is used here.

[0052] Bicubic interpolation first selects a 4×4 interpolation neighborhood on the source image with the target pixel as the center, and constructs a bivariate cubic polynomial model at each pixel in the interpolation area: Among them, p(x,y) is the pixel value obtained after interpolation, a i,j represents the coefficient of the polynomial, x and y represent the relative coordinates of the target pixel in the interpolation area; using the values of the 16 pixels in the interpolation area, 16 binary equations can be obtained, namely: p(x i,j ,y i,j )=f i,j , where f i,j Represents the grayscale value of the pixel in the i-th row and j-th column in the interpolation area, (x i,j ,y i,j ) represents the relative coordinates of the pixel point in the interpolation area; solve these 16 equations to obtain 16 coefficients a i,j , the gray value g of the target pixel is calculated by the obtained coefficient, which can effectively avoid the blurring and distortion problems caused by low-order interpolation methods such as bilinear interpolation and nearest neighbor interpolation.

[0053] Step c: For the features at different levels, the bidirectional feature pyramid module is used to perform feature fusion using top-down and bottom-up paths, and the attention mechanism of the context refinement module is used to enhance the model's attention to the target in the region. First, a convolutional layer is used to generate an initial feature pyramid containing multi-scale features. Then, a bottom-up approach is used to fuse low-resolution features into high-resolution features. Each layer of features is extracted and fused with features before the layer below it, mainly through convolution operations and feature scaling. A top-down approach is used to fuse high-resolution features into low-resolution features to better process small objects and fine location information. Finally, a set of learnable residual connections is used to integrate the information flows in the two directions, and the feature map is iterated multiple times to fuse the bottom-up and top-down information to obtain the final multi-scale feature pyramid, thereby obtaining a better feature representation.

[0054] To improve detection accuracy, the present invention introduces a multi-scale attention mechanism to fuse features of different resolutions and scales, allowing the detection model to detect the target area at multiple resolutions and better capture information about objects of different scales. The present invention combines a bidirectional feature pyramid and a multi-scale attention mechanism.

[0055] Step d: Sample and predict each pixel on the feature map. Use the center branch to determine the position information of the object's bounding box from the coordinates of the key points in the candidate box. Use the regression branch to estimate the size and position of the target box, generate the bounding box of the object, and use a multi-classifier to obtain the probability scores of different target categories in the area. Set a threshold. Target categories with scores below the threshold are considered background, and target categories with scores above the threshold are retained as prediction results.

[0056] On the feature map, each pixel point is taken as the center point of the location area to which it belongs, and the feature information of the point is retained. Multiple sets of rectangular anchor boxes are sampled in each arbitrary direction of the area. Each pixel point is sampled and predicted on the feature map. The position information of the object's bounding box is determined by the coordinates of the key points in the candidate box through the center branch. The size and position of the target box are estimated using a regressor, and the bounding box of the object is generated. A multi-classifier is used to determine whether the area contains an object and determine the bounding box of the object.

[0057] The center branch predicts each pixel one by one, treating each pixel p(i,j) as a candidate detection box, predicting whether the position is the center of the object, and giving a center weight, indicating the distance between the center of the object and the position. The center branch outputs the probability of whether the position is the center of the object. Indicates whether the pixel point is likely to be the center point of an object; if the position is the center point of the object, then its corresponding center score D ij High; otherwise, D ij Low; when performing detection, first select the position with the highest centrality score from the output of the centrality branch as the center point of the detection box;

[0058] The regression branch uses the Smooth L1 regressor to estimate the size and position of the target box and generate the bounding box of the object. The classification branch calculates the probability score of each pixel belonging to each category, adopts a multi-class classifier, and uses the softmax function to calculate the probability score of each category to generate a probability distribution vector, which represents the probability that the feature map belongs to a specific category. The feature map generated by the optical remote sensing ship image to be detected is input into the multi-classification network. The softmax function can convert a K-dimensional vector into K real values between 0 and 1, and the sum of these real values is equal to 1. After normalizing a real vector, a probability distribution vector is generated, x = [x1, x2,…, x k ], xi Represents the score of the i-th category, and the softmax function calculates the probability score of each category: e represents the base of natural logarithm, P i Represents the probability score of the i-th category; the denominator in the formula is the exponential sum of all category scores. Dividing the exponential of the numerator score by the exponential sum gives the probability score of each category.

[0059] Step e: Correct the bounding box using bounding box offset and scaling, then use non-maximum suppression to deduplicate the bounding box. Sorting the bounding boxes from high to low confidence, selecting the bounding box with the highest confidence one by one, and performing an intersection-union calculation on the bounding box with the highest confidence and all the remaining bounding boxes to obtain the accurate object bounding box.

[0060] Since optical remote sensing images contain a lot of information, interference information and duplicate frames will be generated, which need to be removed through NMS. All target frames are sorted in descending order according to their scores, and the target frame with the highest score is taken out and added to the final output result list L, and then removed from the target frame list; the IoU value between the remaining target frames and the target frame with the highest score is calculated, and the target frames with an IoU value greater than a certain threshold are deleted from the list;

[0061] The IoU value between target boxes is used to measure the degree of overlap between two target boxes and is calculated as follows:

[0062]

[0063] Among them, A and B represent target box A and target box B respectively. When the two target boxes completely overlap, their IoU value is 1; when the two target boxes do not have any overlapping parts, their IoU value is 0.

[0064] Step f, calculating the error through the loss function and using the back propagation algorithm to update the model parameters to obtain a small target detection model for ships in any direction;

[0065] In order to prevent the model from overfitting the training data, a regularized cross entropy loss function is used to improve the model's generalization ability and effectively perform classification. The core idea of regularization is to limit the complexity of the model by adding a penalty term to the loss function, thereby reducing the degree of overfitting of the model to the training data. After adding a regular term to the cross entropy loss function, the model will be more inclined to choose a simpler model, thereby avoiding the risk of overfitting. The regularized cross entropy loss function is expressed as:

[0066]

[0067] Among them, L2 represents the regularized cross entropy loss function, y iIndicates the true label of the i-th sample, true is 0, false is 1, γ i Represents the corresponding predicted probability value, N represents the number of samples, / / w / / 2 represents the sum of the squares of the weight vector, λ represents the regularization coefficient, and log represents the natural logarithm. If the detection model performs poorly on the test set, the model parameters can be adjusted based on the actual situation and the value of the regularization coefficient λ can be increased to more smoothly control the complexity of the model and prevent overfitting.

[0068] After obtaining the model error, the gradient descent algorithm based on the chain rule is used to propagate the output error in the reverse direction of the network, calculate the contribution of each parameter to the error, and update the parameters. The specific implementation method is as follows:

[0069] Use the current model parameters to perform a forward propagation on the input to obtain the output result;

[0070] Calculate the loss function, compare the output result with the true label, and calculate the loss function value;

[0071] According to the chain rule, the gradient of each parameter to the loss function is calculated to be the sensitivity of the parameter;

[0072] According to the gradient of the loss function, use a certain learning rate to update the model parameters; repeat the above steps until the value of the loss function converges or reaches a certain number of iterations;

[0073] In the regularized cross entropy loss function, the gradient of the loss function that the backpropagation algorithm needs to calculate consists of two parts: one is the gradient of the cross entropy loss function, and the other is the gradient of the regularization term. For each parameter, the gradient is equal to the partial derivative of the loss function with respect to the parameter plus the partial derivative of the regularization term with respect to the parameter.

[0074] Step g: Use the arbitrary direction ship small target detection model to detect the input remote sensing image.

[0075] First, load the trained target detection model into the memory, then preprocess the image to be detected and convert it into a format suitable for model input. Then input the preprocessed image into the loaded model. The model predicts the image and outputs the detection result, that is, the category and location information of the target object. Finally, the detection result is output to the screen so that the user can intuitively understand the detection result. Figure 4 This is the network structure diagram for ship small target detection. Figure 5 This is a diagram showing the ship target detection effect of the present invention.

[0076] Finally, it should be noted that the above are only specific embodiments of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above specific embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above specific embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes or replacements do not deviate from the spirit and scope of the technical solutions of the specific embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images, characterized in that: The following steps are included: Step a: Using an image slice enhancement algorithm based on a generative adversarial network to perform super-resolution reconstruction on the acquired image; Step b: Adjust the network initial width and width multiplier coefficient to allow the RegNet network to extract input image features, obtain features at different levels, and add deconvolution operations to make the model detect small objects more accurately; Step c: For the features at different levels, the bidirectional feature pyramid module is used to perform feature fusion using top-down and bottom-up paths, and the attention mechanism of the context refinement module is used to enhance the model's attention to the target in the region; Step d: Sample and predict each pixel on the feature map. Use the center branch to determine the position information of the object's bounding box from the coordinates of the key points in the candidate box. Use the regression branch to estimate the size and position of the target box, generate the bounding box of the object, and use a multi-classifier to obtain the probability scores of different target categories in the area. Set a threshold. Target categories with scores below the threshold are considered background, and target categories with scores above the threshold are retained as prediction results. Step e: Correct the bounding box using bounding box offset and scaling, then use non-maximum suppression to deduplicate the bounding box. Sorting the bounding boxes from high to low confidence, selecting the bounding box with the highest confidence one by one, and performing an intersection-union calculation on the bounding box with the highest confidence and all the remaining bounding boxes to obtain the accurate object bounding box. Step f, calculating the error through the loss function and using the back propagation algorithm to update the model parameters to obtain a small target detection model for ships in any direction; Step g: Use the arbitrary direction ship small target detection model to detect the input remote sensing image.

2. The method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images according to claim 1, characterized in that: To process low-resolution remote sensing images containing ship targets, step a includes using an image slice enhancement algorithm based on a generative adversarial network to perform single-frame and multi-frame high-definition reconstruction of blurred targets in the input image and restore the target texture details in the image.

3. The method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images according to claim 1, characterized in that: In order to accurately detect small targets, step b includes: The RegNet model uses the NRB structure to extract features using multi-scale information in each block; A deconvolution layer is added to the feature extraction network to upsample the low-resolution feature map to a high resolution through deconvolution operation.

4. The method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images according to claim 1, characterized in that: In order to reduce information loss and ambiguity, in the bidirectional feature pyramid feature fusion process, in each level of the pyramid, the forward path reduces the feature map of the previous level through convolution and pooling operations, and performs feature fusion with the feature map of the next level; the reverse path enlarges the feature map of the next level through upsampling and convolution operations, and performs feature fusion with the feature map of the previous level.

5. The method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images according to claim 1, characterized in that: The step d comprises: Remap each point in the feature map back to the original image. If the point mapped back to the original image is within the bounding box and the classification category is as expected, then the point is assigned a positive weight, otherwise it is considered negative. The prediction of each pixel is completed by three branches. For each pixel, a square receptive field area is generated with it as the center on the feature map. Each area has a fixed scale and corresponding weight. The classification branch outputs the probability of which category the pixel belongs to; the center point branch predicts whether the pixel is the center point of the object; finally, the bounding box regression branch predicts the bounding box of the object corresponding to the pixel; these three branches jointly process each pixel in each level and ultimately obtain all possible object boxes.

6. The method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images according to claim 5, characterized in that: The original bounding box is corrected by the bounding box offset and scaling predicted by the network to obtain a more accurate object bounding box, and then the non-maximum suppression method is used to remove duplicate detection boxes.

7. A method for detecting small ship targets based on pixel-by-pixel processing of remote sensing images according to claim 1 or 6, characterized in that: In order to prevent the model from overfitting, the classification loss, regression loss and center loss are combined. The loss function of the entire network is to use the cross entropy loss function with L2 regularization to calculate the error, and the back propagation algorithm is used to update the model parameters to obtain the small target detection model of ships in any direction. The cross entropy loss function with L2 regularization is expressed as: Among them, L2 represents the regularized cross entropy loss function, y i Indicates the true label of the i-th sample, true is 0, false is 1, γ i Represents the corresponding predicted probability value, N represents the number of samples, / / w / / 2 represents the sum of the squares of the weight vector, λ represents the regularization coefficient, and log represents the natural logarithm.

Images