Research And Application Of Infrared Image Analysis Based On Deep Learning

With the extensive use of infrared imaging technology in various realms,infrared image analysis has gradually become an important research topic.How to analyze infrared image quickly and accurately is the keynote of research.However,traditional infrared image analysis methods cannot meet the requirements of infrared image analysis.Although the mushroom growth of deep learning in last several years provides a powerful tool to solve the problem of infrared image analysis,the application of deep learning method to infrared image mainly faces two problems:On the one hand,deep learning requires a large number of samples to train the model;on the other hand,infrared images have the characteristics of low overall quality,lack of image details,blurred boundaries,concentrated grayscale,and high background similarity.Considering that the most basic task in image analysis in real scenes is recognition and positioning(ie,target detection),this paper first studies the infrared image target detection task and applies it to infrared image analysis in traffic scenarios.Due to the lack of feature information in the infrared image and the limitation of data set,the detection results show that different types of vehicles are divided into the same category.which is not conducive to subsequent analysis.Therefore,this paper has studied the infrared image classification task and application.This article mainly does the following work:(1)Infrared image target detection based on improved FPN-FRCNNIn order to solve the problem that infrared images are difficult to detect and infrared image data sets are limited,this paper proposes an improved FPN-FRCNN model.Based on the FPN-FRCNN model that can fuse multi-scale features,the shortcomings of applying the FPN-FRCNN model to infrared image target detection are analyzed and corresponding solutions are proposed.The method of transfer learning is used to train the model to solve the problem of limited infrared image data set.By enhancing the feature extraction capability of the convolutional backbone network,the problems of insufficient infrared image quality and lack of detailed information are solved.Rol Align pooling is used to improve the positioning accuracy of the model to solve the problem of poor fuzzy contrast of the infrared image boundary and not obvious texture features.Through Soft-NMS,the rate of miss detection of the model is reduced,and the problem of miss detection caused by target overlapping in infrared image is solved.Finally,the effectiveness of the proposed method is verified by several comparative experiments.(2)Infrared image classification based on improved prototypical networkIn order to solve the problem that deep learning requires many of image samples and infrared image is difficult to classify,an improved prototypical network model is proposed in this paper.Based on the prototypical network,analyze the shortcomings of applying the prototypical network to the classification of infrared images and propose corresponding solutions.The prototypical network can perform few shot learning,the problem of image classification with insufficient samples can be solved.The residual module is used to enhance the feature extraction capability of the feature embedding module to solve the problems of insufficient infrared image quality and lack of detailed information.Change the way of obtaining prototypes,and get more discriminating prototypes as the prototype representation of each category to solve the problem that infrared image features are not obvious.The improved triplet loss is introduced to optimize the metric space,and solve the problems of infrared image gray concentration,high background similarity,and the insufficient difference between classes in the feature space.Finally,the effectiveness of the proposed method is verified by several comparative experiments.(3)Design and implementation of infrared image analysis systemAccording to the method proposed in this paper and the analysis of actual needs,the infrared image analysis system is initially designed,and then the infrared image analysis system based on deep learning is realized through programming.The system is mainly composed of infrared image target detection and infrared image classification.Finally,the system was tested to show the practical value of the system.