Research On Defocus Blur Detection By Fusing Multi-scale Deep Features

When the object exceeds the depths of field of camera lens,the light cannot converge,and then the defocus blur was caused in images.The DBD(defocus blur detection)results can be used in the image quality assessment,blur magnification,saliency detection,image deblurring,etc.We find that the accuracy and robustness of existing DBD methods can also be improved.This paper mainly researches on the DBD method based on multi-scale deep features fusion,and systematically summarizes the traditional or deep learning DBD methods.We proposed two DBD models by researching and analysis of current DBD methods.Our works includes:Research and summary of current DBD methods.According to the feature types used by the current DBD methods,we divide them into two categories: DBD methods based on hand-crafted features and DBD methods based on deep learning.Analyzing and comparing the advantages and disadvantages of these two types of methods,to certain the current problems of DBD.In order to solve these problems and improve the current DBD method,our method was proposed in this paper.Multi-scale blur feature fusion based DBD.Since the scale ambiguity of blur and the current methods are not well performance in specific image areas,like homogenous region.We proposed a multi-scale blur feature fusion(Ms BFF)based DBD model.This model can be divided into two sub-networks: first,multi-scale feature extraction sub-network;second,multi-scale blur map estimation sub-network.The first sub-network extract multi-level and multi-scale fused blur features from input images of different sizes;The second sub-network uses these features to estimate blur map until all fused features are integrated in top-to-bottom manner.After this,the model output the final DBD result.Multi-scale blur map refinement based DBD.Aiming at the shortcomings of the Ms BFF based model,we further propose a multi-scale blur map refinement(Ms BMR)based DBD model.This model also adopts modular design and can be divided into three sub-networks:first,multi-scale feature extraction sub-network;second,multi-scale blur map refinement sub-network;third,multi-scale blur map fusion sub-network.The first sub-network extract two types of features: multi-scale features and fusion features from the input image;the second sub-network infers the multi-scale blur map from the two types of features and refine it.Finally,the third sub-network fuse the multi-scale blur map to obtain the final output result.The above two models both use a two-stage training strategy to ensure that the model weights are optimal.During training,a multi-component loss function and a multi-layer supervision strategy are used to improve model performance.Our model achieved high performance in the DUT and CUHK datasets.