Hello everyone, I am Weixue AI. Today I will introduce to you the deep learning practical 64-black and white photo coloring model application, and quickly deploy the function to quickly color black and white pictures. Image colorization is a challenging problem with multi-modal uncertainty and high ill-posedness. Directly training deep neural networks often results in incorrect semantic colors and low color richness. Although Transformer-based methods can provide better results, they often rely on hand-designed prior knowledge, have poor generalization capabilities, and introduce color bleeding effects. To address these issues, we propose DDColor, an end-to-end image colorization method with dual decoders. Our approach consists of a pixel decoder and a query-based color decoder. The former restores the spatial resolution of images, while the latter exploits rich visual features to refine color queries, thus avoiding the need for hand-designed prior knowledge. Our two decoders work together to establish correlations between color and multi-scale semantic representations through cross-attention, significantly mitigating color bleeding effects. Furthermore, we introduce a simple yet effective color richness loss to enhance color richness. Extensive experiments prove that DDColor outperforms existing state-of-the-art work both quantitatively and qualitatively.
Code and model download address: https://github.com/piddnad/DDColor
1. Black and white photo coloring project
Many old photographs and images preserved in historical archives exist in black and white. These images cannot directly represent the true color of the scene, limiting people's perception and feelings of the past. Therefore, in order to restore the color of these images, researchers conducted research on black and white photo colorization. Colorizing black and white photos is a very challenging task because black and white images both lose color information and lack clear semantics about objects and scenes. Directly applying deep neural networks to black and white photo colorization often results in inaccurate color reproduction and low color richness. Traditional rule-based methods require a lot of manual intervention and manually designed prior knowledge, and this method is difficult to adapt to different situations.