Research And Implementation Of Clothing Extraction Based On Superpixel Segmentation

With the rapid development of the Internet information technology and e-commerce. online shopping is becoming a convenient. efficient and attractive shopping way. getting billions of web users’ attention. Among them. clothing shopping plays a very important role. Usually. web users. especially female customers. painstakingly spend much time each day to browse. search and select desired clothes. Therefore. an effective service for clothing search by visual similarity would have exceptional commercial value. However. the clothes on e-commerce websites usually are taken with natural outdoor background. In addition, these pictures are dressed by fashion models to show and attract customers. These properties make clothing visual search becomes a challenging task. This paper mainly aims at clothing shopping images, and extracting clothing from these images to improve the quality of visual image search. The main contents are as follows:Firstly, a kind of clothing segmentation algorithm combined pose detection and region-based contrast is proposed. This algorithm utilizes superpixel segmentation to segment the image into a series of regions. Then, it uses pose detection to detect the coarse clothing area, integrating the result of image segmentation to localize the seed region. In the next step, spatial information and HSV color feature are combined to measure the region similarity. What’s more, we construct an objective function with the sum of weighted squared error. Therefore, the clothing segmentation problem is converted to minimize the objective function by iteratively computing and reassigning regions. Experimental comparison results show that the proposed algorithm is fast and scalable, allowing our clothing segmentation automatic and effective.Secondly, we propose a clothing extract optimization algorithm based on clothing properties and body structure. For region-based contrast method, the foreground area may contain several unconnected parts. This paper utilizes size and spatial location of the region to construct importance model. Taking into account the integrity of clothing, the pixels surrounded by clothing regions should belong to the clothing. Thus, we assign the interior pixels to foreground class. What’s more, we propose a torso model based on maximum a posteriori, and hence the non-clothing objects such as head and lower limb are further optimized by torso position. At last, GrabCut algorithm is used to refine the clothing extraction result in pixel level. Experiments demonstrate that the proposed algorithm improves the performance of clothing extraction, and can be extended to more clothing Image.