Wavelet Approximation-Aware Residual Network for Single Image Deraining.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Deep convolutional neural networks (CNNs) have advanced single image deraining by learning direct mappings.
• Existing methods struggle to separate rain from object edges and background, leading to loss of detail.
• Complex CNN architectures do not always guarantee effective rain removal and detail reconstruction.

Purpose of the Study:

• To propose a novel wavelet approximation-aware residual network (WAAR) for improved single image deraining.
• To effectively remove rain from both low-frequency structures and high-frequency details.
• To enhance the recovery of image edge details and texture structures.

Main Methods:

• Wavelet transform to decompose images into low-frequency and high-frequency components.
• Novel approximation aware mechanism (AAM) and approximation level blending (ALB) for low-frequency sub-image processing.
• Block connection in high-frequency networks for rain streak elimination and edge enhancement.

Main Results:

• WAAR effectively removes rain while reconstructing clean, rain-free images.
• The method excels at recovering undistorted texture structures and enhancing image edges.
• Experimental results show superior performance compared to state-of-the-art approaches on synthetic and real datasets.

Conclusions:

• The proposed WAAR network demonstrates significant improvements in single image deraining.
• The wavelet-based approach effectively handles rain removal and detail preservation.
• WAAR shows particular strength in recovering image edges and texture details.