Memory Augment is All You Need for Image Restoration

Xiaofeng Zhang, Xuhang Chen, Chaochen Gu, Shanying Zhu, Kim-Fung Tsang, Xinping Guan

Abstract: The quality of images captured by consumer electronic devices, such as smartphones and digital cameras, is often compromised by adverse environmental conditions, leading to degradations like rain, shadows, blur, and low light. These issues significantly impact the user experience. This paper introduces MemoryNet, a novel image restoration framework designed to enhance the quality of images captured with consumer devices. We also propose Degradation-Aware CLIP (DA-CLIP) for the perceptual classification of degraded images, a common challenge in consumer photography. MemoryNet utilizes a three-granularity memory layer and contrastive learning to effectively restore images. The memory layer retains deep image features, while contrastive learning ensures the alignment of learned features for a balanced restoration. We have tested our model on several challenging tasks, including de-raining, de-shadowing, deblurring, and low-light enhancement. The results show significant improvements in PSNR and SSIM in four datasets, demonstrating that MemoryNet can produce restored images with high perceptual authenticity, making it a promising solution to improve the imaging capabilities of consumer electronics. The code can be accessed in https://github.com/zhangbaijin/MemoryNet.

External IDs:doi:10.1109/tce.2026.3655769