Research On Deep Learning And Probabilistic Graphical Model Based Intrinsic Decomposition Method

In recent years,remarkable progress has been made in image understanding,especially in image classification,object detection and localization.However,there is still a lack of effective ways to understand more complex information,such as the shape and material of the objects and illumination.The images formed only by the shape and illumination or only by the material belongs to human visual perception,which are usually called the intrinsic images.Understanding albedo determined only by surface material and shading determined only by shape and illumination for a single image is a longstanding and challenging task.Intrinsic decomposition is to decompose an image into albedo and shading,which is an underconstrained ill-posed problem.It is difficult to decompose the albedo and the shading accurately,due to the lack of sufficient information in a single image.When an image contains a complex scene,the decomposed images often have serious ambiguities among albedo and shading.To solve these problems,this paper proposes a method of intrinsic decomposition based on deep learning and probabilistic graphical model,which can decompose a single image without specific assumptions on material,shape and illumination types.In this paper,a convolutional neural network(CNN)is used to decompose an image into albedo and shading.Then a conditional random field(CRF)is used to optimize the albedo and shading.The proposed CNN model is designed on multi-scale,deep supervision,coarse-to-fine architecture with multi-stage training method to obtain initial intrinsic images,which are much better than other methods.After that,CRF is used to optimize the intrinsic images with initial images and their gradient images inputted,to get detailed and clear boundary intrinsic images.In addition,the proposed CNN and CRF based intrinsic decomposition method was found generality and potential on solving some classical visual problems.