Research On Near Infrared Image Colorization Technology Based On Deep Learning

As an effective means to obtain scene information under low illumination,nearinfrared imaging is widely used in the fields such as monitoring,remote sensing and mine observation.However,the near-infrared image does not contain color information,and has the disadvantages of blurred edges and insufficient details,its visual effect is obviously insufficient compared with the visible color image.As a research hotspot in the field of computer vision and image processing,image colorization aims to give reasonable color to each pixel in the image without color information.At present,gray image colorization has achieved good results,but there is still less research on near-infrared image colorization,and the effect is relatively poor,which can easily lead to the problems of texture blur and the mismatch of texture and color.With the rapid development of deep learning,the research of image colorization using neural network has gradually become the mainstream.Therefore,based on the study of a large number of advanced colorization algorithms at home and abroad,taking the Conditional Generative Adversarial Network as the basic framework,this paper improves the generator and discriminator respectively,and innovatively puts forward a method of using large pixel camera to enhance colorization.The specific research contents are as follows:(1)A texture and color synchronous reconstruction algorithm based on multi-task learning is proposed.Based on the multi-task learning theory,this paper proposes a texture and color synchronous reconstruction model with dual decoder U-net network as the generator,and exchanges features between the two decoders through the Feature Share Module to maintain the consistency of texture and color reconstruction results.Besides,Through the Global Feature Module and Feature Fusion Module,the model integrates global semantic information into the skip-connection between the encoder and the decoder,so as to realize the multi-scale feature fusion and utilization.(2)A near-infrared image reconstruction loss based on double discriminator of autoencoder structure is proposed.The discriminator reconstructs the near-infrared image from the color image by learning the mapping opposite to the generator.Then,the similarity of error distribution between reconstructed image and real near-infrared image is used to calculate the loss according to the statistical characteristics of normal distribution,and the dual discriminator architecture is used to alleviate the problem of mode collapse.(3)An enhanced colorization algorithm based on color guidance of large pixel camera is proposed.Because of the high sensitivity of large pixel camera in low illumination environment,an image acquisition platform coordinated by main and auxiliary sensors is constructed with large pixel camera,color camera and near-infrared camera,and a small dataset is constructed.In order to integrate the color information of large pixel image into the generated image.Based on K-means clustering algorithm,the model generates a palette with five main colors of large pixel image and integrates the palette information into the color reconstruction process.Experiments show that the proposed methods can effectively reduce the blur of colored images and improve the detail and color performance.