Research On Infrared Super Resolution Imaging And Small Target Classification Technology

In recent years,infrared imaging technology has been widely used in automatic driving,intelligent security and remote sensing,but there are still some shortcomings such as low spatial resolution,weak contrast and low signal-to-noise ratio.Superresolution reconstruction technology based on deep learning is an effective method to improve the resolution of infrared image.Based on the practical application background and the characteristics of infrared imaging,the technology and application of infrared super-resolution imaging are deeply studied from the aspects of network structure optimization,subjective and objective imaging quality evaluation and small target classification.The main research work and achievement are as follows:1.Research on super resolution theory and the main structure of neural network.On the basis of CNN and GAN,a multi-scale dense residual block structure and multilevel feature fusion method are proposed to optimize the super-resolution reconstruction network of single infrared image.Compared with interpolation and SRCNN,this method improves the super-resolution reconstruction effect of infrared image,and the PSNR of super-resolution infrared image increases from 28.65 to 31.53.2.In the application of infrared super-resolution imaging,it is actual demand to improve the subjective visual effects of images.At present,most of the image superresolution reconstruction methods based on deep learning are trained and optimized with objective evaluation index as loss function.Therefore,this paper focuses on the correlation between subjective evaluation and quantifiable objective evaluation indexes,and finds that the characteristics of phase consistency are highly correlated with subjective evaluation results.Based on this,a loss function based on subjective and objective joint evaluation is designed and applied to the super-resolution reconstruction algorithm of infrared image.Experiments show that this method can improve the subjective visual effect of image while maintaining the objective evaluation score.3.Detection and recognition of small infrared targets is a major difficulty in the field of infrared detection.In this paper,the use of super-resolution reconstruction technology to enrich the details of targets and improve the accuracy of small target classification is studied.A classification algorithm of small infrared targets based on super-resolution reconstruction is proposed.Super-resolution reconstruction method based on multi-scale residual blocks is used to improve the resolution of infrared small target image and enrich the image feature information.On this basis,convolution neural network is used to classify the enlarged small target image.The experimental results show that the super-resolution reconstruction algorithm in this paper can more accurately extract image features and obtain higher quality of the reconstructed image.Using this method to classify small infrared targets can improve the classification accuracy.In addition,this paper also combines the super-resolution network and classification network to achieve the end-to-end classification of small infrared targets by utilizing the features of infrared images to be reconstructed in the process of superresolution reconstruction.4.Infrared imaging systems have real-time requirements.In this paper,a framework of infrared super-resolution imaging system is designed,and real-time algorithms are studied using embedded digital computing platform.NVIDIA TX1 embedded platform is adopted,which integrates ARM processor and GPU computing unit.In order to further improve the real-time performance,the infrared superresolution reconstruction network model is simplified by pruning and quantization compression,which reduces the computational complexity on the basis of ensuring the reconstruction performance.Finally,the experimental prototype can output high quality and high resolution infrared images in real time.