Deep Direction-Context-Inspiration Network for Defocus Region Detection in Natural Images

Defocus region detection (DRD) problem aims to assign per-pixel predictions of focus clear areas and defocus blur areas. One of the challenges in this problem is to accurately detect the boundary of the transition region between the focus and defocus regions. To address this issue, the paper proposes a direction-context-inspiration network (DCINet), which can take advantage of the directional context effectively. First, we extract directional context by recurrent neural networks initialized with the identity matrix (IRNN) to weight the feature maps and integrate them in the two-group integration method, which can produce the coarse DRD maps. Second, the maps are level-integrated with the source image guiding and the coarse maps are refined gradually. The overall DCINet can integrate low-level details and high-level semantics efficiently. The Experimental results demonstrate that the network can detect the boundary of the transition region precisely, achieving the state-of-the-art performance.