Cosine metric supervised deep hashing with balanced similarity

Deep supervised hashing takes prominent advantages of low storage cost, high computational efficiency and good retrieval performance, which draws attention in the field large-scale image retrieval. However, similarity-preserving, quantization errors imbalanced data are still great challenges deep hashing. This paper proposes a pairwise similarity-preserving scheme to handle aforementioned problems unified framework, termed as Cosine Metric Supervised Hashing with Balanced Similarity (BCMDH). BCMDH integrates contrastive cosine similarity distance entropy preserve original semantic distribution reduce simultaneously. Furthermore, weighted measure metric is designed impact data, adaptively assigns weights according sample attributes (pos/neg easy/hard) embedding process similarity-preserving. The experimental results on four widely-used datasets demonstrate that proposed method capable generating hash codes quality improve performance.