Efficient Compression Blind Artifact Removal Using Prompt-Enhanced Residual Networks

Image artifact removal is a crucial method for enhancing the quality of reconstructed images. By training models to accommodate varying Quantization Parameter (QP) qualities, this technique effectively reduces costs and facilitates practical applications, attracting significant scholarly interest. Current blind Compressed Image Restoration (CIR) studies typically rely on quality factor prediction networks to aid in restoring compressed images. However, these predicted quality factors lack spatial information, limiting the network's adaptability to image content. Recent research on prompt-based image restoration has demonstrated the strong generalization capability of prompts in addressing various types and degrees of degradation. For instance, the PromptIR model can handle multiple functions such as dehazing and deraining, providing inspiration for this study. Consequently, we developed a prompt-based blind image artifact removal model, Prompt-Enhanced Residual Networks (PERN), to improve the restoration of compressed images. To further enhance image quality, we incorporated a Residual Network branch. Experimental results show that our model achieves a 27.63% performance improvement compared to traditional JPEG images.