Image Retrieval with Hard Negative Centers and Quadruplet Center Loss in Deep Hashing

In this work, we enhance image retrieval performance by building upon the Central Similarity Quantization method in deep learning, which uses 'hash centers'—points in a K-dimensional Hamming space fulfilling certain distance criteria. Our approach is inspired by the concept of selecting hard negatives in deep learning, aiming to boost model performance. We introduce an alternative strategy for quadruplet loss training that employs corresponding positive/negative hash centers instead of positive/negative samples. For each sample, we compute the Hamming distances between its hash code and all hash centers and select the two closest centers as negative centers. The positive center is its random assigned center and is used to guide the sample towards convergence. We then define a quadruplet loss function that encourages the learned hash codes to be both discriminative and compact, minimizing the distances between the sample and its positive center while maximizing the distances between the sample and the negative centers. Our method is evaluated using mAP and Recall@K metrics on the Hotels50k, Stanford Dogs, and CUB200 datasets, demonstrating promising results in image retrieval tasks by effectively utilizing hard negative centers and quadruplet center loss.