Saliency Based Blur Measurement And Segmentation For Partially Blurred Images

The partial blur may hinder people to obtain the desired information.Therefore,effective detection and segmentation of these blur regions is very important.Therefore,this paper studies the blur detection and segmentation of the partially blurred image.Our work can be summarized as follows:(1)Saliency based image local blur measurementHow to detect the blurred regions in images with flat textures effectively and rapidly is a known difficulty.This thesis presents a method based on saliency and extreme point.We combine the feature singular value vector of the transform domain,local extreme points that reflect the high frequency information of the image with HiFST(High-frequency multiscale Fusion and Sort Transform)coefficients,train them with BP(Back Propagation)neural network to obtain the blur measure.Then the preliminary results are refined with the salient features of images to obtain the further detection results.Finally,the final results are obtained by optimizing the edge information through bilateral filtering.The experimental results of the precision-recall(RP)ratio on a large dataset show that the proposed method based on saliency constraints is effective especially for images with flat texture.(2)Saliency-based blur region segmentationThe existing methods of blur region segmentation often have a lot of false positives.In this thesis a new energy-based segmentation method for blur regions is formulated,where a new constraint using saliency is included.In the method,adjacent pixels are constrained by the saliency and color similarity.Experimental results show the proposed method is effective for the segmentation of partially blur.