Scholarlink

In this chapter, we propose a two-step process for automatically colorizing grayscale images with high fidelity and explainability. First, grayscale images are converted to color images using an encoder–decoder model combined with a PatchGAN discriminator. Skip connections help preserve local details, while the adversarial mechanism contributes to the sharpness and naturalness of the colors. Next, we integrate the Grad-CAM technique to visualize the model’s decision-making process. Using gradients of two chroma channels, a and b, as weights, we generate heatmaps that highlight the image regions to which the network pays special attention during the colorization process. Experimental results in the COCO-Stuff and DIV2K datasets demonstrate that the method achieves competitive performance in terms of PSNR, SSIM, and DISTS, while preserving transparent textures and edges, as indicated by qualitative evaluation. The built-in explanation module not only sheds light on the internal working mechanism of the network but also lays the foundation for the development of interactive and customizable colorization systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable Grayscale Image Colorization via U-Net-GAN and Gradient-Based Attention

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nhu-Tai Do,
  • Tinh Vu Duc,
  • Quoc-Huy Nguyen

摘要

In this chapter, we propose a two-step process for automatically colorizing grayscale images with high fidelity and explainability. First, grayscale images are converted to color images using an encoder–decoder model combined with a PatchGAN discriminator. Skip connections help preserve local details, while the adversarial mechanism contributes to the sharpness and naturalness of the colors. Next, we integrate the Grad-CAM technique to visualize the model’s decision-making process. Using gradients of two chroma channels, a and b, as weights, we generate heatmaps that highlight the image regions to which the network pays special attention during the colorization process. Experimental results in the COCO-Stuff and DIV2K datasets demonstrate that the method achieves competitive performance in terms of PSNR, SSIM, and DISTS, while preserving transparent textures and edges, as indicated by qualitative evaluation. The built-in explanation module not only sheds light on the internal working mechanism of the network but also lays the foundation for the development of interactive and customizable colorization systems.