| Shadow detection and removal technology is an important groundwork in the computer vision and image processing field and can effectively realize scene recovery,target recognition and feature extraction.It is widely used in agriculture,meteorology,communication,intelligent transportation system,remote sensing and medical diagnosis.At present,the studies on shadow detection and removal have gotten some achievements.However,shadow environment is featured with randomness,diversity and complexity.Also,shadow detection and removal algorithm has certain pertinence,so it is difficult for shadow detection and removal method to satisfy different demands of the application environment.Therefore,we study the shadow detection and removal algorithms deeply,and optimize their algorithms and improve general applicability of the algorithms.The work has import-ant practical significance.The shadow detection and removal issues of a static image are studied and discussed in this dissertation.The purposes and significance of the image shadow detection and removal methods are summarized and the categories and properties of the shadow regions in images are described.Also,we sort out all kinds of shadow detection and removal methods proposed in recent years and study some problems and application limitations of these methods.By aiming at the above problems,three shadow detection and removal algorithms are proposed to overcome some shortcomings of the existing algorithms.Main contents are stated as follows:Firstly,by aiming at the limitation of many automatic shadow detection algorithms,inaccurate image shadow detection for complicated scenes and shadow of multiple discontinuous shadow regions and inaccurate differentiation between the shadow regions and dark region,a new interactive shadow detection method based on Gaussian mixed model is proposed.The interactive method can artificially provide the context information for the shadow detection and reinforce the comprehension on scene contents.First of all,affinity propagation algorithm is used for pixels clustering and downsampling patch image is formed in accordance with the pixel color and position feature,so as to improve the efficiency of detecting the shadow regions.Also,a Gaussian Mixed Model brush containing six-dimensional eigenvector is constructed,including light intensity density,color and texture features.Man-machine interaction can be used to identify the shadow regions and non-shadow region.At last,multi-scale bilateral filter GMM is conducted for images.This can reduce noise in the shadow regions,smoothen the shadow regions and improve the detection precision of the shadow region,while reserving boundary information of the shadow regions.The experimental results prove that the method proposed in this paper has the higher shadow detection precision.Moreover,under the complicated scenes with multiple shadow regions,the relatively ideal shadow detection results can be obtained through many times of interaction and iterative operation.Secondly,according to the linear affinity model used by illumination transfer to shadow removal method,the shortcomings of the global unified transfer operator are used to propose the illumination transfer shadow removal algorithm matched with the self-adaptive sub-regions to realize illumination recovery for the image shadow regions with the complicated texture and uneven illumination.This method adaptively divides the input images into several shadow sub-regions and non-shadow sub-regions in accordance with the distribution of the image shadow locations,image colors and textural features.Moreover,color and textural features in the non-shadow regions are calculated and the sub-regions with the higher similarity in the shadow sub-regions are matched.On the basis of matching sub-regions,the self-adaptive illumination recovery operator is constructed to remove the shadow in the shadow regions.Illumination of adjacent shadow sub-regions is optimized to improve quality of shadow removal results.At last,due to sharp changes of light intensity at the shadow boundary,the shadow boundary regions are conducted the single detection and treatment.This shadow removal method is simple and effective.It can conduct shadow removal for shadow images with the complicated texture structure and images with uneven illumination in the shadow regions.The self-adaptive illumination recovery operator guarantees basic illumination of recovery and maintains harmony and consistency with surroundings in vision.Thirdly,according to the advantages of sparse representation and dictionary learning theory in image recovery,a shadow removal algorithm based on image features and dictionary learning is proposed.First of all,initial image shadow removal is done,thus the shadow regions and non-shadow regions can effectively guarantee precision of regional matching in the similar illumination radiation grade.The non-local sparse model is adopted to construct the association of image blocks in the shadow sub-regions and image blocks in non-shadow sub-regions in accordance with the matching relationship between shadow sub-regions and non-shadow sub-regions with self-adaptive decom-position.Several matched sample blocks are applied to build the group matrix.With the application of image information in shadow regions and the matched non-shadow region information,the image restoring method based on sparse representation and dictionary learning is used for shadow removal to realize illumination unification in regions after removal shadow and non-shadow regions.Finally,by aiming at serious texture detail losses in the dark shadow regions,the morphological component analysis method is adopted to decompose the images,so as to gain the smoothness layer and texture layer of images and conduct the effective compensation for texture information in shadow regions.This method can naturally realize restoration of shadow boundary,thus boundary will be smoother after removing shadow to integrate with the non-shadow regions and meet the subjective visual requirements. |
