[Paper] Product Region Extraction Using DNN-Based Patch Similarity Measurement and Saliency Detection

This study proposes an effective handcrafted algorithm of product region extraction, which identifies product regions at image pixel level from images uploaded to consumer-to-consumer online marketplaces. The proposed method applies graph cuts to classify the image pixels and patches as either product or background regions. A comparative experiment using 425 product images demonstrated that the proposed method significantly improved the F-measure by at least 1.83% compared to a conventional method, with at least 0.13% and 2.16% of increase in precision and recall, respectively. In addition, the proposed method achieved the F-measure comparable to that of a supervised segmentation using deep neural networks, which were trained on 641 product images. Further comparisons and discussions in this paper provide a detailed evaluation of the effectiveness of the proposed method.