Research On The Fusion Of Multi-features Shadow Detection And Light Compensation Shadow Removal Algorithm

Image processing and computer vision are widely applied into various domains.The existence of shadow in images will cause a series of problems,such as object deformation and the loss of feature information etc.The negative effect is inevitable for object detection and object tracking.Based on the dataset which contains street,ground and pedestrian in natural light scenes,this thesis tries to tackle the task of shadow detection and removal.Utilized by the theories of digital image processing and machine learning,we propose a novel adaptive direction tracking filter to achieve shadow detection,which is based on the fusion of multi-feature.Then we study the shadow removal method which relies on the techniques of neighbor region matching and light compensation.The approach we employed in this thesis detect and remove shadows,effectively.The main studies are listed below.In the aspect of pre-process shadow images,since guided filter has the advantage from both image smoothing and image sharpening,we conduct edge-preserving smooth on all the images from dataset by guided filter.For different color spaces represent different feature information,we transform the images after the guided filter process into gray,LAB and ILL color space,which can make us easy to obtain different features in the future study.Based on above operations,we utilize Canny edge detection algorithm to detect the edge of object and shadow in images.An algorithm of shadow detection based on the multi-feature fusion and adaptive direction tracking filter is proposed.In order to solve the problem of low accuracy of shadow boundary detection,this thesis integrates 7 different features into a new feature vector to represent the feature of shadow and non-shadow area.The new vector includes light intensity,different color space and high order statistic features etc.In order to improve the accuracy of shadow detection,we propose an adaptive direction tracking filter which is based on the Canny algorithm.The advantage of this proposed filter is that it can automatically change the window(convolution kernel)according to the direction of shadow or object boundary.This trait enhances the effectiveness of the extracted features.We employ MLP as classifier to mark the shadow boundary.Based on lots of experiments and the comparison with other methods,our method shows better results.The research on shadow removal is based on neighbor region matching and light compensation.To preserve as much as information from original images,we combine the watershed method and region growth method to compute the shadow and nonshadow areas.We employ neighbor region matching algorithm to match shadow region with non-shadow region.Finally,we provide light information from non-shadow region to shadow region,by utilizing light compensation algorithm.This method can remove shadow effectively.Through many simulations,our method shows better results compared with others.