Research On Image Colorization Method Based On Image Generatio

Image coloring refers to the process of filling in missing color information in monochromatic photos or videos.At present,the coloring of grayscale images is still the core task of animated films,medical image processing,and various computer vision applications.At present,many deep learning based image coloring methods have been proposed,such as convolutional neural network(CNN)based methods and generative adversarial network(GAN)based methods.Although these methods have made some progress,there are still some unresolved challenges,such as color monotony in the output image,color overflow,and missing object textures.In response to these issues,this article proposes a new image coloring method based on GAN.The main content and key innovation points are as follows:(1)This article proposes a new dual stream GAN framework(BS-GAN)for grayscale image coloring,which utilizes two parallel encoding and decoding subnetworks to improve coloring performance while learning low-frequency and high-frequency features and preserving image texture details.In order to selectively fuse the features of dual stream sub networks and enhance the importance of channels and positions potentially beneficial for image coloring,this paper proposes a Bistream Feature Extraction Module(BSFEM)module and adds a mixed attention structure to it,which can effectively integrate the Add Pixel Attention(APA)module and Channel Attention(CA)module.To further fuse features and restore chromaticity information,this article also proposes a Feature Boosting Module(FBM)module to combine features of different scales to achieve feature enhancement.In addition,this article adopts the Multi scale Feature Attention Module(MSFAM)module,which distinguishes and determines the colors of different regions through multiple Global Context Block(GCB)blocks,thereby further enhancing the texture details of restored color images.(2)This article proposes a dual Patch GAN discriminator network based on Markov discriminator(Patch GAN).Unlike ordinary discriminators,dual Patch GAN discriminators not only focus on the authenticity of global and local information,but also attach importance to the analysis of local patches in objects,conducting small-scale dense discrimination and analysis,and promoting the generator to generate more realistic and edge clear color images.(3)This paper proposes a new combined loss function,which uses both the color loss function and WGAN(Wasserstein GAN with Gradient Penalty)loss function.The combination of two functions can stabilize the training of the network while making the generated color image chromaticity close to the real image,and generate detailed color images.The experimental results show that the proposed method can effectively solve the problems of color overflow and missing texture information in image coloring,and generate vivid and realistic color images from grayscale images.Meanwhile,compared with other mainstream image coloring methods at present,our method has achieved the best performance in both qualitative and quantitative comparison.On the Image Net dataset,compared with the most advanced methods at present,the peak signal-to-noise ratio(PSNR)index of our method has increased by 18%,and the structural similarity(SSIM)index has increased by 8%.