This work was co-authored by Zonghe Shao, Qichao Wang, Yuzhe Cao, Yijin Gong, Zhuodong Luo, advised by Prof. Hao Lu.
Defocus Blur Detection aims to separate in-focus and out-of-focus regions from a single image pixel-wisely.
Image segmentation? Obviously not!
We proposed PRNet, based on Encoder-Decoder framework. In the Encoder, ResNet18 is used for multi-scale image feature extraction and Patch Attention Module(PAM) is used to perform local to global attention analysis at different scales.The Decoder consist of embedded Residual Learning and Refinement Module(RLRM), which allows the top-down and bottom-up feature fusion and decodings.
DUT-DBD dataset: Defocus Blur Detection via Multi-Stream Bottom-Top-Bottom Network
CUHK dataset: Discriminative Blur Detection Features
Comparison with existing work.
Performance in some extreme scenarios.
