Defocus Blur Detection via Multi-Stream Bottom-Top-Bottom Network

Wenda Zhao  ,Fan Zhao  , Dong Wang , Huchuan Lu

1 Dalian University of Technology, China
2 Dalian Institute of Chemical Physics, Chinese Academy of Sciences, China

Kindly Note

This paper is accepted by TPAMI, which is an improved version for MSBTBNet in CVPR2018, e.g., (1) a new CRLNet is designed to gradually refine the blur detection maps through cascaded DBD map residual learning from the small scale to the large scale; (2) we extend our DBD dataset is extended by adding 600 challenging images with pixel-level annotations and making the explicit training-testing protocol. Thus, our dataset (DUT-DBD) consists of 600 training images with pixel-level annotations and 500 testing images with pixel-level annotations.

The code and DUT-DBD dataset can be downloaded below.


Fig. 1 The pipeline of our DBD algorithm. Each colorful box is considered as a feature block. The arrows between blocks indicate the information stream. Given an input image, its multi-scale versions generated by the resize operation are first encoded in the bottom-top stream by a modified VGG16 model [28], respectively. Then, the integration of bottom-top and top-bottom streams is performed by feedback and forward information combination modules (FFICs). After that, CRLNet is designed to gradually refine the preceding blur detection maps from the small scale to the original scale through learning the residual.


Defocus blur detection (DBD) is aimed to estimate the probability of each pixel being in-focus or out-of-focus. This process has been paid considerable attention due to its remarkable potential applications. Accurate differentiation of homogeneous regions and detection of low-contrast focal regions, as well as suppression of background clutter, are challenges associated with DBD. To address these issues, we propose a multi-stream bottom-top-bottom fully convolutional network (BTBNet), which is the first attempt to develop an end-to-end deep network to solve the DBD problems. First, we develop a fully convolutional BTBNet to gradually integrate nearby feature levels of bottom to top and top to bottom. Then, considering that the degree of defocus blur is sensitive to scales, we propose multi-stream BTBNets that handle input images with different scales to improve the performance of DBD. Finally, a cascaded DBD map residual learning architecture is designed to gradually restore finer structures from the small scale to the large scale. To promote further study and evaluation of the DBD models, we construct a new database of 1100 challenging images and their pixel-wise defocus blur annotations. Experimental results on the existing and our new datasets demonstrate that the proposed method achieves significantly better performance than other state-of-the-art algorithms.


Code and Datasets

Code for our method is available here.

Our dataset (DUT-DBD) is available here.

Results on Shi’s dataset and DUT-DBD for our method are available here.