In this paper, we propose a kansei retrieval method for living room images so that users can intuitively retrieve their desired properties on real estate portal sites. The proposed method employs a two-stages training model that combines Conv-DCCAE (Deep Canonically Correlated Auto-Encoders) and ranking model. For training model, a set of anchor images, positive examples of kansei words that represent impressions of the anchor images, and negative examples of kansei words that differ from the impression of the anchor images are input. First, the relationship between kansei words and images is trained using Conv-DCCAE, and projected onto the common space. Then, a ranking model is trained, and a kansei retrieval space is constructed by correcting the distance of positive examples to the anchors to be smaller and the distance of negative examples to the anchors to be larger in the common space. The proposed method also personalize the kansei retrieval space by retraining the ranking model using relevance feedback. The retrieval performance was evaluated by using nDCG (normalized Discounted Cumulative Gain), which is an index to evaluate the performance of the ranking. The proposed method achieves the most accurate ranking for nDCG@1, 5, 10 compared with the conventional methods. In addition, the proposed method was able to provide more appropriate retrieval results to users, whose initial retrieval results were inappropriate, by using relevance feedback.
View full abstract