Cross-domain Clothing Image Retrieval Based on Hybrid Shape Feature

Nowadays, Islamic woman garment attracts many users, indicated by the increasing amount of garment industry in many countries. Most of garment purchasing is done electronically. Keywords are inputted to the online shopping system to choose the expected garment model. However, it is very complicated to describe a garment model using keywords. Sketch-based image retrieval (SBIR) is an approach of retrieving a garment based on the content of the clothing. The objective of this paper is to enhance Pyramid Histogram of Oriented Gradient (PHOG) performance on SBIR for Islamic woman garment. This research combined the PHOG and Moore-Neighbour tracing (MNT) method for extracting a new feature. The oriented gradient of the image is calculated; the MNT was used to detect the boundary of the image. Then, the hybrid feature of the oriented gradient and the image boundary is created. The spatial pyramid is then generated from the feature. We refer the proposed method as pyramid histogram of oriented gradient cross-feature (PHOGXF). We conducted simulations on 450 image datasets, which consisted of three categories: blouse-pant, long-dress, and tunic. We did a cross-domain retrieval; we retrieved the clothing images in the database using a sketch image as the query. The result showed that the proposed method obtained the precision value of 73.5% for blouse-pant, 90.4% for long-dress, and 52.1% for the tunic, while the PHOG only achieved the precision value of 18.0% for blouse-pant, 65.4% for long-dress, and 21.0% for the tunic. We concluded that the proposed method (PHOGXF) has significant performance; more than twice as good as the PHOG method.