Research On Grayscale Image Coloring Algorithm Based On Deep Learning

Image coloring has always been a hot topic in computer vision.Image coloring is a task that designing algorithm to apply color values to the pixels of grayscale images,to make them into color images,enrich the information contained in the image and enhance its perception.Grayscale images always inevitably appear in all fields of real life,so image coloring has a wide range of applications,including but not limited to artistic creation,medical imaging,security surveillance,and even military fields.Since the advent of deep learning and excellent performance in multiple image processing tasks,convolutional neural network(CNN)-based image coloring methods have emerged one after another and demonstrated relatively excellent results,and the study of image colorization has set off a boom again.The related research work is generally divided into single-image coloring method and multi-image coloring method.In view of the shortcomings of the two types of methods,two deep learning coloring methods are proposed in this paper to help the development of image shading technology.The specific research work is as follows:(1)This paper first introduces the relevant research of image shading,introduces the shortcomings of traditional shading algorithms and user-interaction algorithms in practical applications,and introduces two types of methods for generating color images independently by network models.(2)Using U-Net as the backbone network to perform basic feature extraction and color reconstruction,and inspired by the commonality of image coloring tasks and image semantic segmentation tasks,an end-to-end network model is designed to enhance semantic understanding by feature fusion,and a single-image coloring method based on full-scale feature skip connection is proposed.The channel attention module is introduced to further accurately obtain the feature channels related to the shading task,highlighting the key points while avoiding noise interference.After experimental tests,the proposed algorithm can make up for the common problems in the current image coloring work,and produce rich details and more realistic color images.(3)A multi-image coloring method based on bilateral attention mechanism and attention enhancement convolution is proposed,specifically using a coloring algorithm that introduces color reference images,the algorithm uses the unique nature of attention mechanism to establish long-distance connections between features to achieve implicit color propagation,the network extracts the feature information of grayscale images and reference images through two encoder subnets,designs a feature guidance module(FGM)to output the fused features,and then strengthens the guidance effect through attention enhancement convolution.Then,the decoder network restores the features and rebuilds the colors.Finally,the generative adversarial network is used to guide the network to generate finer and more vivid images.The algorithm can quickly capture the color of the reference feature,accurately transfer it to the target feature,and effectively solve the problem of small target lost during the downsampling process.Through experimental verification,the results show that the method is effective,surpassing many of the latest image coloring algorithms in human testing,visual effects,and evaluation indicators.