Defocus Blur Detection Based On Convolutional Neural Network

Defocus blur detection is aimed to detect defocus blur region and focus clear region of the input image,which is a pixel-level task and has been highly applyed in automatic focusing,image restoration and other computer vision fields.Recently,deep convolution neural network has shown its powerful feature extraction ability in defocus blur detection task and has made great progress.However,most methods of convolution neural network always rely on costly pixel-level annotations.In order to reduce the annotation cost,this thesis proposes to accomplish defocus blur detection only with the box-level annotations.Box-level annotations can provide important cues about the approximate position of the defocus regions and the focus regions,but lose the boundary of the transition regions in detail.To solve this problem,we introduce a recurrent constraint network after full convolution neural network,which generates pixel-level results through static training based on box-level supervision,and then the results are used as annotations to fine-tune the network to optimize new pixel-level annotations for dynamic training.We iterate continuously between generating pixel-level results and fine-tuning pixel-level annotations.To boost the performance further,a guided conditional random field is developed to improve the annotation quality,and the annotations are corrected at the same time.In order to promote the further research of weak supervision methods,this thesis constructs and publics a new dataset named FocusBox,which contains 5000 challenging images and box-level annotations.In addition,in order to solve the problems of inaccurate boundary in transition region and black hole noise in homogeneous region,we explore the interaction between region and edge detection of defocus blur.Hollowing the region results can reduce the edge discontinuity,and filling the edge results can reduce the black hole noise.The hollowing-filling interactive learning can effectively improve blur detection results.Based on this point,we design three kinds of interactive structures,including series network,parallel network and series-parallel hybrid network.In summary,this thesis proposes two methods in weak supervision field and full supervision field to complete the defocus blur detection based on convolution neural network.Experimental results verify that our method with weak supervision not only yields comparable results than fully supervised counterparts but also achieves a faster speed,and the proposed hollowing-filling interactive learning with full supervision can significantly improve the detection results.