FFAFR: Feature Fusion Attention Network for Colorful Filter Removal

Social media images are often added filters for aesthetic and other reasons. However, these filters can cause information perturbation and color imbalance in pixels. This can make it difficult for neural networks to interpret the original and filtered versions as the same image, which significantly affects the performance of downstream visual tasks. To tackle this issue, we propose a Feature Fusion Attention Network for Colorful Filter Removal called FFAFR, with a Global Channel Attention Block (GCABlock) for feature extraction, and a novel Mixed-scale Feature Fusion Module (MFFM). The experiments demonstrate that FFAFR can effectively and efficiently remove image colour filters. On the IFFI dataset, FFAFR outperforms the previous state-of-the-art method, achieving a 0.71dB improvement in PSNR and demonstrating better qualitative results.