Research On Image Colorization Method Based On Deep Learning

Image colorization aims to assign appropriate colors to each pixel in a grayscale image to generate a visually richer colored image.This technique has wide application prospects in digital art,medical imaging,and film production.However,grayscale images can correspond to multiple colors at the same gray level,which makes the image colorization problem uncertain.Generating reasonable colored images is a challenging task in the field of computer vision.With the improvement of computer image processing capabilities and the continuous progress of technology,deep learning-based image colorization research has also made progress in the efficiency and color effect of generating images.However,there are still some problems,such as a lack of detail information and color overflow.Therefore,based on the review and analysis of domestic and foreign research status,this thesis proposes some improved automatic coloring algorithm,with the main research contents and innovations as follows:(1)To address the problem of a lack of detail information,a colorization method combining attention mechanism and instance-awareness is proposed.On the premise of reducing the amount of manual operation,attention mechanism and instance-awareness are combined to improve the accuracy and rationality of coloring in image details.The model effectively locates and learns meaningful object-level semantics by using the instance-aware coloring algorithm.At the same time,some attention mechanism module is introduced in the coloring stage to enhance the performance of coloring detail information.The model uses object detection technology to obtain cropped objects and then uses two similar image coloring network branches,one for extracting object-level features and the other for extracting whole-image-level features.Finally,the fusion module is used to fuse the whole-image-level and object-level features to predict the final image colors.Experimental results on the COCO dataset show that after combining the SENet,CBAM,and ECA visual attention mechanism models,the PSNR is improved by 4.46%,3.63%,and 3.77%,and the SSIM is improved by0.74%,0.53%,and 0.60%,respectively.(2)To address the problem of color overflow,a colorization method based on coupled luminance and chrominance total variation model regularization is proposed,which introduces the regularization term of the coupled luminance and chrominance total variation model into the model.Convolutional neural networks can be used in image colorization tasks to calculate possible colors for each pixel and their corresponding probability distribution.However,the final selection does not consider the continuity and regularity of the image,resulting in color overflow.To improve this,a coupled luminance and chrominance channel total variation model is designed as the regularization term of the model to restrict the adjustment direction of the model parameters,to maintain the consistency between the output color and the boundary of the object in the grayscale image,and to avoid color overflow.Experimental results on the COCO dataset show that after the model introduces the regularization term of the coupled luminance and chrominance total variation model,the PSNR is improved by 9.94%,and the SSIM is improved by 1.42%.(3)To simultaneously address the issues of lacking fine detail information and color overflow in the coloring effect,above two methods are combined by introducing the regularization term of the coupled luminance and chrominance total variation model into the model structure of the image colorization method that combines attention mechanism and instance-awareness.Experimental results on the COCO dataset show that after the two methods are combined,the PSNR is improved by 13.41%,and the SSIM is improved by 2.47%.