A feature library migration remote sensing small sample target detection method

Abstract

The application discloses a remote sensing small sample target detection method for feature library migration. Firstly, multi-scale feature extraction is performed on the input remote sensing image to be detected based on a deep convolutional neural network to generate a detection model of common categories; a small sample feature library is set, and the small sample feature library is added to the detection model for training, so that the detection model can maintain the detection ability for common targets while detecting small samples; in view of the high symmetry of the remote sensing image to be detected, a symmetric matrix parameter coefficient is introduced when the detection result is predicted after the training of the detection model is completed, the positioning branch parameter is corrected, and the small sample target detection precision is improved. The above method solves the problem that the modeling of the target category features is poor in the remote sensing image, and the target cannot be accurately recognized, and improves the effectiveness of the remote sensing image detection in actual application.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to a feature library migration remote sensing small sample target detection method. BACKGROUND

[0002] The target detection problem is an important problem in the field of remote sensing image processing, and the detection of small sample targets is a more difficult task scenario, but has very high application value. Small sample target detection is a task of accurately positioning and identifying target categories that are not common or have a small number of samples. Small sample target detection technology learns a dataset composed of common category samples and a small number of rare category samples, extracts feature information of the target from the dataset, and constructs a model that can accurately identify and locate common category targets and relatively accurately identify and locate rare category targets.

[0003] In recent years, small sample detection technology has developed rapidly, and there are many excellent small sample target detection algorithms. These algorithms can achieve good performance for targets with a large scale in the image (the length-width ratio of the image and the target is generally not more than 10) and a small number of targets under the condition of small samples. However, for remote sensing images, due to their large image size, large image-to-target length-width ratio, and wide range of target sizes, existing target detection technology has poor performance when applied to small sample detection tasks with only a small number of samples. The detection model cannot accurately detect and identify the target through a small number of training samples. SUMMARY

[0004] The purpose of the present application is to provide a feature library migration remote sensing small sample target detection method, which solves the problem of poor feature modeling for a small number of target categories in remote sensing images and cannot accurately identify, and improves the effectiveness of remote sensing image detection in practical applications.

[0005] The purpose of the present application is achieved by the following technical solutions:

[0006] A feature library migration remote sensing small sample target detection method, the method comprising:

[0007] Step 1, first, for the input remote sensing image to be detected, multi-scale feature extraction is performed based on a deep convolutional neural network to generate a detection model for non-small sample categories;

[0008] Step 2, setting a small sample feature library, and adding the small sample feature library to the detection model for training, so that the detection model maintains the detection ability for common targets while detecting small samples;

[0009] Step 3, in view of the characteristics of high symmetry of the remote sensing image to be detected, a symmetric matrix parameter coefficient is introduced when the detection result is predicted after the training of the detection model is completed, the regression branch parameters output by the detection model are corrected, and the small sample target detection precision is improved.

[0010] It can be seen from the technical solutions provided by the above-mentioned application that the above-mentioned method solves the problem that the modeling of the target category characteristics is poor in the remote sensing image with a small number of target categories, and cannot accurately identify, and improves the effectiveness of the remote sensing image detection in practical application. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0012] Figure 1 The feature library migration remote sensing small sample target detection method flowchart provided by the embodiment of the present application. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0014] As Figure 1 The feature library migration remote sensing small sample target detection method flowchart provided by the embodiment of the present application is shown, the method comprises:

[0015] Step 1, first, based on the deep convolutional neural network, the multi-scale feature extraction is carried out for the input remote sensing image to be detected, and the detection model of non-small sample category is generated;

[0016] In this step, first, the remote sensing data set A containing non-small sample category targets is set, the position coordinates and category labels of the non-small sample category targets are labeled, and it is ensured that there is no small sample target in the remote sensing data set A or the small sample target does not participate in the loss function calculation in the training;

[0017] The convolutional neural network based on FasterRCNN is set, the feature extraction module of the convolutional neural network is set as the high-resolution feature extraction module of HRNet, and the convolutional neural network is trained by using the remote sensing data set A.

[0018] Step 2, set up a small sample feature library, and add the small sample feature library to the detection model for training, so that the detection model can maintain the detection ability of common targets while detecting small samples;

[0019] In this step, a support branch is added to the trained detection model, a remote sensing data set B containing small sample targets is set up, and the small sample targets are labeled with position coordinates and categories to ensure that there are no common targets in the remote sensing data set B or the common targets do not participate in the loss function calculation as negative samples in training;

[0020] Other parameters of the detection model are fixed, and the support branch and the classifier are trained separately to obtain the feature extractor parameters of the small sample targets;

[0021] For each sample of the remote sensing data set B, the sample size is K, and the feature generated by each sample through the support branch is

[0022] The vector is m, the feature vector generated by the sample through the detection model is n, and the average vector of the K samples is The feature library of this category is constructed, the similarity of the two is measured, the cosine similarity is selected, and the similarity β generated by each small sample target is used as the parameter coefficient of the detection model branch;

[0023] The prediction result output of the detection model is divided into a classification branch for distinguishing target categories and a regression branch for positioning target positions; the last layer 1*P vector of the classification branch is removed, P is the number of common targets, Q is the number of small sample targets, and a layer 1*(P+Q) vector is initialized as a classification vector; the model is trained on remote sensing data sets A and B; when training the categories in the remote sensing data set A, the parameter coefficient of the classification branch is 1; when training the categories in the remote sensing data set B, the coefficient of the classification branch is the corresponding similarity β generated in step 4;

[0024] After the detection model converges, the parameter coefficient of the classification branch is set to 1, and the training is performed again until convergence, and the detection model training is finally completed.

[0025] Step 3, in view of the high symmetry of the remote sensing image to be detected, when the detection model is trained, the symmetry matrix parameter coefficient is introduced to modify the regression branch parameters of the detection model prediction result output, and the small sample detection target detection accuracy is improved.

[0026] In this step, specifically, the feature vector of the remote sensing image to be detected is multiplied by multiple symmetric matrices, and the similarity, i.e., the distance measure, of each result to the original vector is compared, and the half of the included angle of the vector with the smallest distance is taken as the target orientation, and a rotation matrix is added to add coefficients to the last layer parameters of the detection model, specifically:

[0027] After removing the support branch, the input remote sensing image to be detected is subjected to feature extraction and feature library vector comparison, and the vectors of the regression branch and the classification branch are obtained;

[0028] For the vector of the regression branch, after decoding the coordinate plane, according to the symmetry in the two-dimensional plane, it is multiplied by the following four matrices respectively, and the cosine similarity of the four feature vectors obtained to the original vector is calculated, and the half of the included angle of the vector with the smallest distance is taken as the target orientation, and the included angle is denoted as θ, and the four matrices are:

[0029]

[0030]

[0031]

[0032]

[0033] According to the included angle θ, a rotation matrix is added to add coefficients to the last layer parameters of the detection model, and the rotation matrix is:

[0034]

[0035] Finally, the classification result is obtained through the softmax classification of the prediction result of the classification branch, and the positioning prediction is performed through the regression branch.

[0036] It is worth noting that the contents not described in detail in the embodiments of the present application belong to the prior art known to those skilled in the art.

[0037] In summary, the method described in the embodiments of the present application can better detect small sample categories and common categories in small sample target detection tasks, and the performance is better than that of existing algorithms, and in remote sensing images, the performance is much better than that of other methods; the present application is aimed at the high symmetry of remote sensing image targets, and by introducing a symmetric matrix transformation, accurate positioning and classification supervision are provided, and the accuracy of remote sensing small target detection is further improved.

[0038] In addition, the method described in the present application is compared with the current mainstream small sample target detection algorithm, and the accuracy in the remote sensing image scene has been greatly improved, and the model detection performance is better than that of the existing algorithm under the condition that the sample amount of each small sample target is single sample, three samples and five samples, which proves the effectiveness of the present application.

[0039] Those skilled in the art can understand that all or part of the steps in the methods of the above embodiments can be instructed by programs to the relevant hardware to complete, and the corresponding programs can be stored in a computer readable storage medium. The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0040] The above description is merely preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements easily conceived by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims. The information disclosed in the background section of the present application is merely intended to deepen the understanding of the overall background of the present application, and should not be regarded as acknowledging or implying in any form that the information constitutes the prior art known by those skilled in the art.

Claims

1. A remote sensing small sample target detection method for feature library migration, characterized in that, The method comprises: Step 1, first, for the input remote sensing image to be detected, multi-scale feature extraction is performed based on a deep convolutional neural network to generate a detection model for non-small sample categories; Step 2, a small sample feature library is set, and the small sample feature library is added to the detection model for training, so that the detection model maintains the detection capability for common targets while detecting small samples; In step 2, a support branch is added to the trained detection model, a remote sensing data set B containing small sample targets is set, the small sample targets are labeled with position coordinates and categories, and it is ensured that there is no common target in the remote sensing data set B or the common target does not participate in the loss function calculation in the training; The other parameters of the detection model are fixed, and the support branch and the classifier are trained separately to obtain the feature extractor parameters of the small sample targets; For each class of samples of the remote sensing dataset B, there are K samples, each sample has a feature vector generated by the support branch m, and a feature vector generated by the detection model n, and the average vectors of the K samples are respectively The category feature library is constructed, the similarity of the two is measured, the cosine similarity is selected, and the similarity β generated by each small sample target is selected as the parameter coefficient of the detection model branch. The prediction result output of the detection model is divided into a classification branch for discriminating target categories and a regression branch for positioning target positions; the last layer 1*P-dimensional vector of the classification branch is removed, P is the number of common targets, Q is the number of small sample targets, a layer of 1*(P+Q) vector is initialized as a classification vector, and the remote sensing data sets A and B are trained; when training the categories in the remote sensing data set A, the parameter coefficient of the classification branch is 1; when training the categories in the remote sensing data set B, the coefficient of the classification branch is 4 corresponding to the similarity β generated in step 1; After the detection model converges, the parameter coefficient of the classification branch is set to 1, and training is performed again until convergence, and finally the detection model training is completed; Step 3, in view of the high symmetry of the remote sensing image to be detected, a symmetry matrix parameter coefficient is introduced when the detection model completes the training and predicts the detection result, the regression branch parameters of the detection model prediction result output are corrected, and the small sample detection target detection precision is improved.

2. The remote sensing small sample target detection method of feature library migration according to claim 1, characterized in that, In step 1, first, a remote sensing data set A containing non-small sample category targets is set, the non-small sample category targets are labeled with position coordinates and categories, and it is ensured that there is no small sample target in the remote sensing data set A or the small sample target does not participate in the loss function calculation in the training; A convolutional neural network based on FasterRCNN is set, the feature extraction module of the convolutional neural network is set as the high-resolution feature extraction module of HRNet, and the convolutional neural network is trained using the remote sensing data set A.

3. The remote sensing small sample target detection method of feature library migration according to claim 1, characterized in that, In step 3, the feature vector of the remote sensing image to be detected is multiplied by multiple symmetry matrices, and then the similarity, i.e. the distance measure, of each result with the original vector is compared, and half of the included angle of the minimum distance vector is taken as the target direction; a rotation matrix is added to add a coefficient to the last layer parameter of the detection model; specifically: After removing the support branch, the feature extraction and feature library vector comparison are performed on the input remote sensing image to be detected, and the vectors of the regression branch and the classification branch are obtained; For the vector of the regression branch, after decoding the coordinate plane, according to the symmetry in the two-dimensional plane, the vector is multiplied by the following four matrices respectively, the cosine similarity of the four feature vectors obtained with the original vector is calculated, half of the included angle of the minimum distance vector is taken as the target direction, the included angle is denoted as θ, and the four matrices are: According to the included angle theta, a rotation matrix is added as a coefficient of the last layer parameter of the detection model, and the rotation matrix is: Finally, the prediction result of the classification branch is subjected to softmax classification to obtain a classification result. And positioning prediction is performed through the regression branch.

Images