Study On The Colorization Of Grayscale Image Based On Convolutional Neural Network

Grayscale image colorization is an important research direction in the field of image processing and has important application value in medical diagnosis,industrial detection and military investigation.Traditional grayscale image colorization method requires human intervention,which leads to the difficulty of realizing colorization efficiently,batch and automatically.With the rapid development of deep learning,especially convolutional neural network,the method of grayscale image colorization based on convolutional neural network has become a hot topic in colorization research.In order to solve the defects of traditional grayscale image colorization method,in this paper we propose a colorization method based on convolutional neural network to realize automatic colorization of grayscale images.In addition,in order to improve the colorization effect and image saturation,the proposed method is optimized to obtain a good colorization effect.Main work of this thesis:1.A colorization method based on convolutional neural network is proposed and the colorization results are analyzed.The results show that the average values of the comprehensive assessment function(CAF),saturation(S)and peak signal-to-noise ratio(PSNR)of the output colorization images are 54.437,114.294 and 25.229 d B,respectively.The average value of S is 1.222 lower than that of the real color images,which basically realizes the colorization of gray image.2.In order to solve the problem of color unsaturation,the previous colorization method is optimized by scene classification.Compared with the previous method,the values of CAF,S and PSNR of the output colorization images are increased by 1.025,1.112 and 0.464 d B,respectively.The colorization effect is further improved and the problem of color unsaturation is solved.3.Based on generative adversarial learning,a new method of grayscale image colorization is proposed and the color saturation is further improved.The results show that the values of CAF and S of the output colorization images are 56.200 and 115.622,respectively.Both of them are infinitely close to the results of the real color images,and the image saturation is significantly improved.