Image Hashing via Linear Discriminant Learning

Deep Hashing with Hash Center Update for Efficient Image Retrieval

“…Another interesting work in a similar direction was proposed by Hong et al in [6]. Here, Linear Discriminant Analysis characteristics are trained directly on the hash codes, thus enforcing the deep network to produce hashes which have a small intra-class variance while also having a high inter-class variance.…”

“…The proposed method updates the hash centers during deep hashing training. Here, the natural problem arises that hash centers are desired to be binary, such that the CNN features are encouraged to be discrete and are desired to be real valued at the same time, such that gradient descent optimization is feasible [6]. Therefore, a distinct treatment of those hash centers is designed depending on whether performing a forward pass or backward pass step during training is implemented.…”

“…Inspired by the idea of updating hash centers during training [6], an alternative method of hash center update based on the weighted mean of the hash values is proposed in our paper, which aims to reflect the movement of the formed clusters in the Hamming space during training. The network was trained using a CCA-based loss formulation such that the correlation between the hash codes and hash centers is maximized along with the correlation of classification scores and class labels.…”

“…The training and test curves, precision-recall curves, and t-SNE [18] plots of the generated feature space were plotted and discussed. Mean average precision was computed and the performance is compared with other supervised hashing approaches such as DPSH [13], DCCH [9], CSQ [21], DSDH [12], DDSH [8], DTSH [19], LDH [6], HashNet [2], DHN [22], DNNH [11], and CNNH [20].…”

“…Data independent approaches [22,17,29] mainly rely upon random projections to generate the hash functions. Locality-sensitive hashing (LSH), is [6] a popular data-independent approach, which was used in [15] and [9] to solve the ANN problem while avoiding the curse of dimensionality inherently associated with exact nearest neighbor searches for data in their original metric spaces. Recent methods mainly concentrate on, datadependent approaches which are mainly categorized as unsupervised and supervised methods.…”