Depth-aware Shadow Detection And Removal From A Single Image

Shadows are natural phenomena in our daily life.They can provide important information for the understanding of visual scenes,such as lighting environment,scene geometry,object shape and other information.However,the existence of shadows will also cause deterioration to the scene analysis,which will make the image processing work such as object recognition,image segmentation and intrinsic decomposition of images worse.Therefore,detecting and removing shadows automatically is an important topic in the field of computer vision.Detecting and removing shadows automatically from a single image is a very challenging problem in the image processing field.Most existing algorithms use the similarities of chrominance,texture and gradient or deep learning technique to achieve shadow detection and removal for images of simple scenes.For complex scenes,due to its diverse structure and complex shadow information,it is difficult to construct a training set,the shadows are difficult to detect and remove ideally by using traditional methods or deep learning methods.Moreover,the illumination condition and geometrical structures of complex scenes are variable,it is a challenging problem to remove the shadows while preserve the original shading information and geometric information,and keep the reality of the image.In order to solve the problem of shadow detection and removal in complex scene automatically,thesis proposes a shadow detection and removal algorithm based on image depth cues.The algorithm uses the depth map of the image to estimate the normal information,point cloud position and the spatial neighborhood corresponding to each pixel.The algorithm estimates the shadow confidence value of each pixel by comparing the similarity between the normal,spatial position and chrominance information between each pixel and its spatial neighborhood pixels,and optimizes the shadow confidence to get the final shadow detection result and solve the difficulty of shadow detection in complex scenes.With the estimated shadow detection result we construct a shadow removal global optimization algorithm,remove the shadow in the scene while restore the shading and chromaticity information in the original image,ensure the reality of the image,and obtain visually friendly shadow removal result.The research content and main contributions of thesis are divided into the following points:(1)Propose a shadow confidence estimation algorithm based on multi-scale filtering to reduce the influence of texture while retaining complex shadow boundary information;(2)Using Laplacian operator to construct an interpolation-based optimization equations,we can obtain better shadow detection results,while retaining the relative intensity of shadow information and shadow boundary gradient information,which helps to remove boundaries of complex shadows and soft shadows;(3)A non-local neighborhood-based shadow removal optimization algorithm is proposed.Based on the assumption that the chrominance information of the image and the ambient illumination information remain unchanged before and after the shadow removal,the chromaticity constraint is proposed and the normal similarity is improved.The calculation method restores the chromaticity and shading information of the shadow elimination result and preserves the authenticity of the result.The experimental results show that the algorithm has better shadow detection and elimination effects for both simple scenes and complex scenes.Finally,the effectiveness and robustness of the proposed algorithm are verified by visual comparison and objective evaluation with the shadow detection and removal results of other algorithms.