Feature Fusion-Based Re-Ranking for Home Textile Image Retrieval

In existing image retrieval algorithms, negative samples often appear at the forefront of retrieval results. To this end, in this paper, we propose a feature fusion-based re-ranking method for home textile image retrieval, which utilizes high-level semantic similarity and low-level texture similarity information of an image and strengthens the feature expression via late fusion. Compared with single-feature re-ranking, the proposed method combines the ranking diversity of multiple features to improve the retrieval accuracy. In our re-ranking process, Markov random walk is used to update the similarity metrics, and we propose local constraint diffusion based on contextual similarity. Finally, the fusion-diffusion algorithm is used to optimize the sorted list via combining multiple similarity metrics. We set up a large-scale home textile image dataset, which contains 89k home textile product images from 12k categories, and evaluate the image retrieval performance of the proposed model with the Recall@k and mAP@K metrics. The experimental results show that the proposed re-ranking method can effectively improve the retrieval results and enhance the performance of home textile image retrieval.