The Research On Virtual Fitting Technology Based On Image

Although the development of apparel e-commerce is rapid in recent years, while users can not accurately estimate the effect of their own fitting only by pictures and documents. This situation decreases the online sales quantity of clothing, and increases higher return rate at the same time, which is harm for the development of online shopping garment industry. To solve this problem, a growing number of researchers have been working on technologies related in virtual fitting system. They built virtual fitting system and provided users with virtual fitting experience.Some achievements have been made of the research in virtual fitting system, but some issues still need further study and rely on the advance of image processing. In this paper, the key algorithm of virtual fitting system is explored, mainly in the following three aspects: generation of a garment library used in virtual fitting, search space reduction and the ability against background disturbance, and the method for adapting different users’ stature and pose.In order to satisfy the need of robustness and automation in clothing matting of image-based virtual fitting system, this paper uses the research achievements in saliency of natural images for reference and takes the saliency of cloth color into consideration. An improved Grab Cut clothing matting method is proposed in this paper, which is constrained by human face detection and the saliency map of global color histogram. An improved Haar cascade classifier is proposed, which based on face detection, upper body detection and eye detection to exclude irrelevant areas. The improved Haar cascade classifier confirms the proper face and upper body area of a model, and enlarges them appropriately to include the cloth region, then the method focuses on the saliency of cloth color histogram, thus we can obtain the mask containing possible foreground pixels. Next, the saliency map and the cloth region estimation window are combined to gain the mask of possible foreground pixels and it is used for the input parameter of Grab Cut. Thus, we can improve the level of automation of clothing matting based on Grab Cut and ability against background disturbance.The improved Haar cascade classifier and the improved Grab Cut clothing matting algorithm based on image saliency also provide convenience for human pose estimation. The estimated region of possible human body reduces the search space. An improved appearance model based on joint attribution and PS(Pictorial Structure) model is used in this paper for upper body pose estimation. With the help of saliency area, the human pose estimation algorithm searches the optimal location points in the areas of skin and cloth and estimates the skeleton of human upper body using body parts’ constraints. Human pose estimation is significant in more accurate clothing matting, virtual fitting and automatic deformation of garment which combines the image warping method based on quadrilateral mesh in the future research.To make virtual fitting more realistic for different users’ statue and pose, image warping method based on quadrilateral mesh is proposed in this paper. It can achieve partial distortion on obtained cloth. Then the cloth is overlaid to user image and adjusted to appropriate location through global transformation. The experiment results show that both the clothing matting method and the image warping method work well in all four combinatorial situations: simple or complex cloth and background, respectively.The results of this paper demonstrate the effectiveness of automation process of clothing matting method, the effectiveness of human body estimation, the effectiveness of cloth warping method and the reality of virtual fitting.