Maximum-likelihood approximate nearest neighbor method in real-time image recognition

作者：

Highlights：

• Maximal-likelihood search improves the performance of the nearest neighbor method.

• Proposed method is 2-10 faster than brute force, randomized kd-tree and perm-sort.

• Unlike the baseline DEM, proposed method does not require quadratic memory space.

• The method is applied with similarity measures which do not met metric properties.

摘要

•Maximal-likelihood search improves the performance of the nearest neighbor method.•Proposed method is 2-10 faster than brute force, randomized kd-tree and perm-sort.•Unlike the baseline DEM, proposed method does not require quadratic memory space.•The method is applied with similarity measures which do not met metric properties.

论文关键词：Approximate nearest neighbor method,Large database,Maximum likelihood,Real-time pattern recognition,Image recognition,Probabilistic neural network,HOG (histograms of oriented gradients),Deep neural networks

论文评审过程：Received 19 August 2015, Revised 20 April 2016, Accepted 15 August 2016, Available online 17 August 2016, Version of Record 30 August 2016.

论文官网地址：https://doi.org/10.1016/j.patcog.2016.08.015