Research On No Reference Blur Region Segmentation Algorithms

Blurry images hide detailed information from real scenes,and it is a harmful phenomenon for many tasks.For many systems,there is an urgent need for a fast and effective blur detection method.On this background,the blur region detection and segmentation algorithms are studied and explored.This kind of algorithms can estimate and generate the dense blur map of the entire image.However,traditional methods based on artificial features have some problems such as low robustness,difficulty in using global information,slow speed,and only a single type of blurry image that can be processed.It is difficult to meet the actual needs.To solve these problems,two algorithms are proposed.To handle the problem of slow speed,a fast blur region segmentation algorithm is presented.In order to solve the problems of traditional methods such as low robustness,a semantic segmentation neural network for blur detection is proposed.The neural network can automatically extract the feature of the blurry area.The main research content of this article is as follows:(1)Based on the blur quality evaluation index of no-reference structural sharpness(NRSS),a fast blur region segmentation algorithm is proposed,which significantly improves the processing speed.The core idea of this method is to down-sample the original image to reduce the computational overhead.And then the sliding window method is used to obtain the blur amount estimation of the entire image.(2)A neural network called BlurNet for blur detection is proposed.The improvements and innovations of this method mainly include: extending FCN network to FCN2 s,adding batch normalization layer before feature combination to make model training faster and more stable,adding dropout layer to prevent over-fitting,exploring the influence of different transfer learning strategies.This method can achieve a better segmentation result without requiring a clear image as a reference and can handle images of different sizes and different types of blur.Compared with other methods,the segmentation results increase significantly.Compared with other semantic segmentation models,such as SegNet and UNet,our method performs better.(3)Using multi-task learning methods,we train the two tasks of blur type classification and blur area segmentation at the same time in one neural network.And two different multi-task learning networks are designed.The performance of these networks is compared and discussed through experiments.