Sparse projections for high-dimensional binary codes

摘要: This paper addresses the problem of learning long binary codes from high-dimensional data. We observe that two key challenges arise while and using codes: (1) lack an effective regularizer for learned mapping (2) high computational cost computing codes. In this paper, we overcome both these problems by introducing a sparsity encouraging reduces number parameters involved in projection operator. not only overfitting but, due to sparse nature matrix, also leads dramatic reduction cost. To evaluate effectiveness our method, analyze its performance on nearest neighbour search, image retrieval classification. Experiments challenging datasets show method better accuracy than dense projections (ITQ [11] LSH [16]) with same code lengths, meanwhile is over order magnitude faster. Furthermore, more accurate faster other recently proposed methods speeding up encoding.