Research on Image Colorization Based on Deep Learning

Abstract Gray scale picture colorization, a research hotspot in the domain of computer vision, seeks to assign reasonable colours to every single pixel of a grayscale picture, enriching the visual information of the coloured picture. Neural networks take grayscale images as input and output colorized images. From a psychological perspective, colours can provide observers with a more pleasant perceptual experience. Old photos can also be coloured, which are important documentary resources for recording real history and restoring the social landscape of the time. Additionally, due to the limitations of imaging mechanisms, most medical images are grayscale. Therefore, research on grayscale image colorization is a task of significant importance. However, existing automatic image colorization systems have issues such as “blurry boundaries”, “colour overflow”, “inappropriate colour choices” and “incorrect colour regions”. This paper proposes a model combining adversarial neural networks and U-Net networks, optimizing them by incorporating global feature fusion modules, Pathway emphasis mechanisms and spatial focus mechanisms. The idea of Markov discriminator was added to the discriminator, and the loss function was optimized. The ultimate investigative conclusions demonstrate that the PSNR and SSIM of the suggested system are 24.45 and 0.941, accordingly, in relation to the preceding protocol, which are 3% and 5% higher than the previous algorithm, and the example has been ameliorated to some degree. And the loss function is modified, and the final algorithm can improve the effect of picture colouring, which encompasses a broad spectrum of application prospects.