Research On Infrared Image True Color Conversion Algorithm Based On Deep Learning

Infrared imaging is highly robust,able to adapt to different weather and lighting conditions,and has broad application prospects in the fields of military,security,and assisted driving.However,the contrast of infrared images is low and lacks color information,which is not conducive to the visual perception of human eyes.Therefore,on the basis of maintaining the advantages of infrared imaging in adapting to weather and light,colorizing infrared images will help improve people’s perception and understanding of infrared images,which has important research significance and practical value.There are obvious deficiencies in the current infrared image colorization methods,most of which are based on generative adversarial networks,and realize unsupervised true color conversion of infrared images by training multiple generators and discriminators.Few infrared visible data sets are disclosed,resulting in a single research scenario.The current infrared image true color conversion methods are converted on low resolution image,can not be applied to high resolution infrared image true color conversion.To solve the above problems,this thesis carried out research on the construction of high-quality infrared visible data set,unsupervised infrared image true color conversion model and method,and high-resolution image conversion,and achieved satisfactory results.The main work of this thesis is as follows:(1)Aiming at the problems that the current infrared data set is small in scale and not suitable for infrared image true color conversion task,a large-scale unpaired infrared visible image data set is constructed in this thesis.Infrared and visible video under urban,high-speed and monitoring scenes were collected by vehicle infrared camera and fixed monitoring equipment,and infrared visible image data was further processed.Finally,the infrared visible image data set containing 6 scenes was constructed,which alleviated the problem of insufficient data set in the research direction of infrared image true color.(2)For unsupervised infrared image true color conversion,the method of contrast learning avoids the use of multiple generators and discriminators,thus reducing the complexity of the model.A backbone network based on residual attention is designed.By introducing channels and spatial attention mechanisms,the network pays more attention to the characteristic channels and spatial locations that carry more useful information,which improves the performance and efficiency of the model.The information transfer between the shallow network and the deep network is enhanced by the residual connection and the network convergence is accelerated.A multi-scale discriminator network is designed to enhance the discriminant ability and improve the image generation level of the generator.The pixel-level constraint is abandoned and spatially dependent loss is used as the structural constraint of the transformation model to achieve cross-domain appearance transformation and effectively maintain the consistency of scene structure.The experimental results show that the proposed method achieves good results on real data sets,and the converted true color images are clearer in detail and more complete in content retention.(3)For high-resolution infrared image true color conversion,and it is difficult to collect a large number of infrared visible images in some scenes,this thesis builds a single-input highresolution infrared image true color conversion model based on pre-trained deep neural network.The pretraining model is used as an external semantic priori to extract the representation of structure and appearance.In order to improve the perception and representation of the model,a generator network was designed to aggregate multi-scale segmentation attention and intensive connections.This method only needs a pair of infrared visible images as input to convert the input infrared image into a true color image.Experiments show that this method can convert infrared images with resolution of more than 2K to true color,and has good scalability.In this thesis,the characteristics of infrared image and the deficiency of relevant research are fully considered,and a variety of technical means are adopted to achieve true color infrared image.Experimental results show that the proposed method can realize unsupervised infrared image true color conversion in multiple scenes,and effectively improve the quality of true color image.In addition,a large scale unpaired infrared visible data set is constructed,which provides strong support for relevant research based on infrared images.