Near Infrared Image Colorization Based On Deep Learning And Multi Resolution Fusion

Image colorization is to transform gray black-and-white image into colorful image.Near infrared image colorization is a branch of image colorization.Compared with the ordinary gray-scale image,the near-infrared image colorization faces the problem of unclear color results due to the limit of original near-infrared image.Near infrared image is widely used in traffic assistant driving,military field,security monitoring,intelligent police,unmanned driving and so on.However,due to the lack of color information,it is not conducive to human observation,so colorizing near infrared image is of great practical significance.With the development of deep learning,some methods show their excellent ability of image colorization.However,there are still many problems in the existing methods of infrared image colorization.However,the existing infrared image colorization methods still have many problems,such as unclear colorization results,abnormal color after coloring,and unstable color regions.In response to existing problems,we propose a color network structure of symmetric skip connection encoder-decoder.The encoder mainly uses the down sampling structure to extract the input image features and understand the image.The decoder mainly uses the upsampling structure to learn the features,recover the image and carry out intelligent coloring.A jumping connection is added between them,which copy the encoder feature maps symmetrically to the decoder in order to reduce information degradation.In order to train the color network better,we innovatively propose the edge aware network to assist the training of the color network,so that the output color region is more stable and the boundary is clearer.Edge aware network is a kind of loss function which is adaptive to color network in training.Its input is a difference map between output and target image of color network.In the process of training,the parameters of color network are optimized by back propagation algorithm to minimize the value of difference map.Aiming at the problem of more or less information degradation caused by too many layers of neural network,the original near-infrared image and the color image output by neural network are fused to retain the highfrequency details in the near-infrared image.We propose a multi-resolution composite fusion method,which can fuse different layers according to different fusion rules and achieve good results.In this paper,a database of near-infrared image and color image pixel level matching is collected and made,and the database under simulated rainy days is obtained by post-processing.The main scene of the database is road image,which can effectively train the network and improve the generalization of the network to achieve the purpose of auxiliary driving.There are1978 pairs of images in each database.In the experimental part,we verified the excellent performance of the overall model on the colorization of near-infrared images through various data sets,and achieved good results.Based on deep learning and traditional image algorithms,this paper makes a progressive study of database establishment and production,color image neural network construction and improvement,and color image fusion enhancement.The overall model of the paper enhances the coloring of near-infrared images to improve human recognition.Intelligent coloring that is difficult to achieve manually can now be automatically colorized by our algorithm model.