Wave goodbye to blurry, noisy, hazy images!
Restoring images by understanding degradation's unique spectral fingerprint.
The paper introduces AdaIR, an adaptive all-in-one image restoration network. It tackles diverse image degradations using frequency domain analysis for effective and efficient restoration within a single model.
-----
Paper - https://arxiv.org/abs/2403.14614
Original Problem 😞:
→ Existing image restoration methods are often specialized for specific degradation types like noise or blur.
→ These methods lack generalizability and require separate models for each degradation.
→ All-in-one models address multiple degradations but often ignore frequency domain information.
→ Different degradations impact different frequency bands, requiring tailored restoration.
-----
Key Insights 🤔:
→ Different image degradations affect distinct frequency subbands.
→ Noisy and rainy images are contaminated with high-frequency content.
→ Low-light and hazy images are dominated by low-frequency degradation.
→ Effective all-in-one restoration should consider these frequency variations.
-----
Solution in this Paper ✨:
→ AdaIR network is proposed for adaptive all-in-one image restoration.
→ It uses Adaptive Frequency Learning Blocks (AFLB) within a Transformer-based U-shaped architecture.
→ AFLB contains Frequency Mining Module (FMiM) and Frequency Modulation Module (FMoM).
→ FMiM extracts low and high-frequency features guided by adaptive spectral decomposition of the degraded input image.
→ FMoM facilitates interaction between these frequency features using bidirectional attention units (H-L and L-H).
→ This allows adaptive restoration by emphasizing informative frequency bands based on input degradation.
-----
Results 📈:
→ AdaIR outperforms PromptIR by 0.63 dB PSNR on average across dehazing, deraining, and denoising tasks.
→ Achieves a 2.27 dB PSNR gain over PromptIR on image deraining.
→ In single-task setting, AdaIR improves PSNR by 0.49 dB on dehazing and 1.86 dB on deraining compared to PromptIR.
→ On five-degradation tasks, AdaIR shows 1.86 dB average PSNR gain over IDR, with over 5 dB gain on dehazing.
Share this post