Co-attention enabled content-based image retrieval

• DOI: 10.1016/j.neunet.2023.04.009
• Publication Date: 2023-04-23T23:51:01Z
• Authors: Zechao Hu; Adrian G. Bors
• Keywords: Benchmark (surveying); Content-Based Image Retrieval; Feature (linguistics); Similarity (geometry)

Content-based image retrieval (CBIR) aims to provide the most similar images a given query. Feature extraction plays an essential role in performance within CBIR pipeline. Current studies would either uniformly extract feature information from input and use it directly or employ some trainable spatial weighting module which is then used for similarity comparison between pairs of query candidate matching images. These modules are normally non-sensitive only based on knowledge learned during training stage. They may focus towards incorrect regions, especially when target not salient surrounded by distractors. This paper proposes efficient sensitive co-attention1 mechanism large-scale tasks. In order reduce extra computation cost required sensitivity co-attention mechanism, proposed method employs clustering selected local features. Experimental results indicate that maps can best benchmark datasets under challenging situations, such as having completely different acquisition conditions its match image.