Wavelet-Enhanced Transformer for High-Fidelity Image Colorization

In the field of image colorization, we propose an enhanced model based on the VQGAN-Transformer framework, incorporating wavelet transforms to better emulate the hierarchical structures of human visual perception and improve the quality of unconditional colorization. While the traditional VQGAN-Transformer framework effectively captures global information in image generation, it still exhibits limitations in detail fidelity and color consistency. To address these issues, we introduce wavelet transforms between the image encoding and quantization stages, leveraging their multi-scale analysis capability to extract hierarchical features, allowing the colorized images to more closely resemble natural human perception. In our method, wavelet transforms are embedded between the image encoder and quantizer, decomposing image features across multiple scales to construct a hierarchy aligned with the human visual system, while re-quantization enhances color and detail fidelity. In summary, this study validates the effectiveness of wavelet transforms in unconditional image colorization, providing a new approach to improving model performance in colorization tasks.