PROJECT TITLE :

Deep Saliency Hashing for Fine-Grained Retrieval

ABSTRACT:

Using hashing algorithms for large-scale web media search has proven to be useful and efficient in recent years However, the current general hashing algorithms lack the ability to distinguish between fine-grained objects that have an overall look but differ in subtle ways. We apply the attention method for the first time into the learning of fine-grained hashing algorithms to tackle this problem. Deep saliency hashing (DSaH) is an unique hashing model that automatically mines salient regions and learns semantic-preserving hashing codes at the same time, as demonstrated in this paper. End-to-end, attention and hashing networks make up the DSaH model. These losses are the semantic loss, the saliency and quantization losses of our function. Discriminative areas are mined from pairs of images using an attention network that is guided by the loss of saliency. Extensive studies are carried out on both fine-grained and general retrieval data in order to evaluate performance. A comparison of our DSaH's fine-grained retrieval performance to that of the strongest competitor (DTQ) on both Stanford Dog and CUB Bird shows that our DSaH is superior to the strongest competitor (DTQ) by about 10%. There are various state-of-the art hashing methods that DSaH is comparable to, such as CIFAR-10 and NUS-WIDE.