Research On Key Technologies Of Infrared Aerial Target Recognition And Detection With Few Shots

Infrared imaging has the advantages of long detection distance,high concealment,penetrating smoke and all-weather day and night work.It has received extensive attention at research and application in the field of electronic reconnaissance.In the infrared detection system,target recognition and localization can provide category and coordinates of potential targets in received images,which is not only the basis for tracking tasks,but also a strong support for system decision-making.In recent years,with improvement of computing power and big data,deep learning has made extraordinary progress in computer vision.Some models have even exceeded human discrimination on many large datasets.However,the performance of these algorithms depends heavily on the adequacy of training data.In the recognition of ground-to-air infrared aircraft tasks,the cost of collecting sample data is very expensive.Some aircraft types are so rare that we only have few labeled images of these aircrafts,and it is difficult to establish a sufficient data library for all types of infrared aircraft types.This paper focuses on the problem of infrared aerial target recognition with few samples,and combines image processing,pattern recognition,sparse representation,and deep learning.The main work and innovations are as follows.?1?An infrared strip noise correction method based on sparse representation is proposed.Firstly,train a dictionary with clean infrared images by K-SVD dictionary learning algorithm,which may improve the expression ability of the output dictionary.Then compute the sparse representation coefficients of the noise image by OMP algorithm,and reconstruct the noise image with dictionary and sparse representation coefficients.Afterwards,obtain the correction coefficients line by line with least square method.Finally,output correction image by noise image and correction coefficients.The experimental results show that the proposed method has a relatively stable correction effect on strip noise,and is applicable to any scenarios,meanwhile has a high tolerance for sparse error tolerance.?2?An infrared aerial target recognition model based on discriminant sparse representation is proposed.Firstly,compute the target main orientation by a novel method,then rotated the target to a reference direction.Secondly,an over-complete dictionary is learned from histogram of oriented gradient features of these rotated targets.Thirdly,a sparse representation model is introduced and the identification problem is converted to an ₁l-minimization problem.Finally,different aircraft types are predicted based on an evaluation index,which is called residual error.Experimental results show that the proposed model has strong rotation invariance,discrimination ability and anti-noise performance.?3?In order to solve the problem that the amount of sample data is seriously insufficient,a few-shot infrared aircraft classification algorithm based on improved relation network is proposed.In this method,the relation network is improved by combination with multi-scale feature and meta leaning.Firstly,a multi-scale feature extraction module is constructed to extract the feature information of input images,then the feature tensors of the support samples and test samples are input into relation module,and different aircraft types are predicted by relation value.The results of experiments show that the proposed method can realize the recognition tasks with only few samples.?4?Aiming at the small size and sparse distribution of infrared aircrafts,a few-shot infrared aircraft detection method is proposed.This method improves the feature reweighting model to enhance the attention of small size targets.In feature extraction module,the output feature map is enhanced by the fusion of deep and shallow convolution features.In reweighting module,instead of the binary image input as the label information,a heatmap is used to enhance the association between infrared aircrafts and their neighborhood,besides,it helps the model to focus on smaller targets.These two modules,together with a prediction module,are trained on an episodic few-shot learning scheme.Experimental results show that the proposed method can achieve detection task with only few labeled samples.