Research And Implementation On Infrared Image Super-Resolution Algorithm Based On Generative Adversarial Network

With the rapid development of deep learning,image processing technology based on deep learning has been widely used in various fields.Among them,passive infrared imaging technology,as a typical representative of imaging technology,has been playing an important role in military defense,medical treatment,forest fire prevention,security and other aspects.Due to the production scale and production cost,it is still difficult to obtain infrared image with resolution comparable to visible image.Therefore,using low-resolution infrared camera and image super-resolution algorithm is a relatively feasible method to obtain low-cost and highresolution infrared images.In view of the current problems in the research and application of infrared image super-resolution,this thesis has carried out research in the following aspects.In order to solve the problem of weak expression ability of neural network model under shallow network,this thesis designs and applies the FB(Feedback Block)referring to the feedback circuit in analog circuit.The output data of the FB is fed back to the input end of the module through the gain coefficient,so that the FB is not sensitive to the noise of the input data,and the anti-interference ability of the module is improved.At the same time,due to the use of the self-feedback mechanism,the FB is easy to enter the oscillation state.In the training stage of the neural network,the gain coefficient is modified to adjust the state of the network to ensure the stability of neural network model and FB.Aiming at the problem of poor super-resolution infrared images obtained by traditional methods,the SRFBGAN(Super-resolution Feedback Block Generative Adversarial Network)based on FB is constructed in this thesis,which has powerful image generation ability to complete image reconstruction.The perceptual loss function is used to improve the perceptual effect of the restored image,and the wavelet coefficients as loss function is used to improve the ability of the model to recover high frequency details.At the same time,infrared imaging technology imaging through thermal radiation,compared with the visible image,the content of infrared image is fuzzy,so visible images are used to construct data sets during training process of SRFBGAN to improve the ability of network model to recover image details.Knowledge distillation is used to optimize the structure of SRFBGAN to solve the deployment problem on embedded platform.According to the characteristics of a large number of flat regions in the infrared image,the smoothness degree of the image content is classified,and three different scale student networks are constructed according to the smoothness degree of the image content.Finally,the restored image blocks will be spliced,which can achieve faster processing speed and ensure better infrared image recovery effect.Finally,this thesis designed an infrared image super-resolution camera platform,selected the hardware of the embedded platform from the engineering point of view,and designed the embedded platform software system.At the same time,in order to ensure the expansibility of the project development,the expansion interface is reserved in the software and hardware,and the human-computer interaction system of infrared image super-resolution analysis camera is developed to provide services for other applications based on infrared image.