Clothing Recognition And Parsing Based On Coocurence Information And Perceptual Grouping

With the booming of electronic commerce, there are increasing needs for applications like virtual dressing room, recommended systems for clothes and similar garments retrieval systems. Clothing recognition and clothing parsing, which implement images and pixels level annotation correspondingly, are the foundations of those applications. The complexity of the backgrounds, the variaties of poses, clothes styles in the images and the limitation of labeled datasets makes it hard for clothing regcognition as well as parsing, and thus result in undesirable performances of the majority existing models. Therefore, in this paper, the relative characteristics of clothing categories and the main factors which affect the performance of clothing parsing models are further studied. Several improved algorithms for clothing recognition and parsing are proposed.The main research work of this paper includes:(1)Image level clothing recognition:To recognize clothes is to label clothes images with exact categories automatically. The mainstream algorithms for clothes recognition are based on single task training models which ignore the coorcurrence relations among the clothes categories. Eventhough some of the existing models consider these relations, such models are hardly conductive to the following features selection. Aiming at this problem, this article incorporate the clothes cooccurrence term into the objective function of multi task learning to fit the relations among clothes categories. Experimental results show that the average performance of the proposed model outperforms the one of single task learning, neural network and traditional multi-task learning.(2)Pixel level clothing parsing:To parse clothes is to label pixels/regions in clothes images with exact categories automatically. First of all, even the majority of existing clothing parsing models take the symmetricity of human bodies into account, however, those models neglect the symmetricity of the clothes themselves; Besides, it is a common problem for some of the existing models that the perdicted labels amongs the boundaries of different clothes regions are ambiguous and easily labeled incorrectly. In order to solve these problems, we modify the structural conditional random fields model (CRFs) in refer to the concept of perceptual grouping from the domain of perception. To be more specific, a binary potential function which embodies the symmetry of the clothes themselves and encourage regions with similar symmetric structures to share the same label. Furthermore, according to the theory of contour closure, we incorporate a cost function of contours gap into the basic loss function to lower the error arounds the boundaries of the predicted regions. This cost term penalizes the predicted clothes regions boundaries which didn’t align to the ground truth. Finally, based on the experimental result of the proposed clothes recognition models, the sigle task logistic learning are replaced by the multi task learning model for the sake of higher performance. The experimental results show that the average performance of the proposed model is higher than the basic clothes parsing models based on structural CRFs.