Research And Implementation Of Clothing Extraction Combining Salient Object Detection With Image Segmentation

The rapid development of the Internet information technology and e-commerce makes the convenient and efficient online shopping becoming an indispensible shopping way. There are a large number of clothing commodities with various types on the network. How to effectively retrieve the target product is one of the most important factors determining whether or not the consumer will purchase and businesses will profit. Thus the related product retrieval technology has become a hot research topic in academia and industry. Some existing content-based image retrieval services cannot achieve good performance because it directly extracts overall vision features such as color, which will be influenced by the noises of cluttered background. In order to improve the quality of the clothing image search, it is necessary to wipe off cluttered backgrounds and extract the clothing. This thesis mainly aims at clothing shopping images with complex backgrounds, and the clothing extraction problem is studied and discussed. The main contents are as follows:Firstly, a clothing saliency analysis algorithm combined with pose detection is proposed. It converts the clothing location to the salient object detection. Human models in the clothing images and the apparel’s visual saliency characteristics provide relevant cues to locate clothing’s approximate area. So the algorithm is built on the basis of pose detection and salient object detection approaches. First, the model’s pose in the image is estimated; then, it was fused to calculate the region salient features in the salient object detection; finally, join local and global saliency maps to get their linear combination to obtain a clothing saliency map combined with pose detection. Experimental comparison results show that the proposed algorithm have effective analysis on clothing saliency, which will help the subsequent preliminary location of clothing region.Secondly, a new clothing extraction algorithm for fashion apparel photographs is proposed. Firstly, the algorithm utilizes the proposed clothing saliency analysis approach in this thesis to obtain the rough region of clothing; secondly, the image is processed by superpixel segmentation, and the color distribution of foreground/background is represented on the superpixel level by Gaussian Mixture Model; then, related image segmentation techniques are exploited to extract clothing; finally, combining saliency detection of clothing object, preliminary extraction results are refined, and hence the non-clothing objects such as hands, head and neck, etc. are further removed at the same time acquiring the image only containing clothing. Experiments demonstrate that the proposed algorithm is well able to combine salient object detection and image segmentation to effectively extract clothing, and can be extended to larger image datasets.