Hashing method maps similar data to binary hashcodes with smaller hammingdistance, and it has received a broad attention due to its low storage cost andfast retrieval speed. However, the existing limitations make the presentalgorithms difficult to deal with large-scale datasets: (1) discreteconstraints are involved in the learning of the hash function; (2) pairwise ortriplet similarity is adopted to generate efficient hashcodes, resulting bothtime and space complexity are greater than O(n^2). To address these issues, wepropose a novel discrete supervised hash learning framework which can bescalable to large-scale datasets. First, the discrete learning procedure isdecomposed into a binary classifier learning scheme and binary codes learningscheme, which makes the learning procedure more efficient. Second, we adopt theAsymmetric Low-rank Matrix Factorization and propose the Fast Clustering-basedBatch Coordinate Descent method, such that the time and space complexity isreduced to O(n). The proposed framework also provides a flexible paradigm toincorporate with arbitrary hash function, including deep neural networks andkernel methods. Experiments on large-scale datasets demonstrate that theproposed method is superior or comparable with state-of-the-art hashingalgorithms.
展开▼
