HHF: Hashing-guided Hinge Function for Deep Hashing Retrieval

Deep hashing has shown promising performance in large-scale image retrieval. The process utilizes Neural Networks (DNNs) to embed images into compact continuous latent codes, then map them binary codes by function for efficient Recent approaches perform metric loss and quantization supervise the two procedures that cluster samples with same categories alleviate semantic information after binarization end-to-end training framework. However, we observe incompatible conflict optimal positions are not identical ideal hash because of different objectives terms, which lead severe ambiguity error-hashing process. To address problem, borrow Theory Minimum-Distance Bounds Binary Linear Codes design inflection point depends on bit length category numbers thereby propose Hashing-guided Hinge Function (HHF) explicitly enforce termination prevent negative pairs unlimited alienated. Such modification is proven effective essential training, contributes proper intra- inter-distances clusters better accurate retrieval simultaneously. Extensive experiments CIFAR-10, CIFAR-100, ImageNet, MS-COCO justify HHF consistently outperforms existing techniques robust flexible transplant other methods. Code available at https://github.com/JerryXu0129/HHFhttps://github.com/JerryXu0129/HHF.