Cross-propagation parallel network for reflection image inpainting

The development of deep learning has led to great advances in image reflection removal techniques. Most existing image reflection removal methods can produce reasonable results but still suffer from missing structure and blurry textures. The main reasons are the following facts: (1) Separation-based methods struggle to distinguish fine structures due to the similarity between reflection and background layers. (2) U-Net may cause the reintroduction of reflection features. To tackle above issues, a novel Cross-Propagation Parallel Network (CPPN) is proposed for reflection image inpainting. Firstly, the correct information from the reflection-free regions of the blended image is used to infer information from the reflective regions based on image inpainting mechanisms to improve the accuracy of structural features. Secondly, a cross-propagation parallel network is designed to further constrain structure and detail for accurate feature representation and propagation. Experimental results on several public datasets show that the proposed method can produce higher-quality results than state-of-the-art methods.