Efficient Image Classification Based on Virtual Subwindows and Random Ferns

摘要

The limitations of migrating image classification methods that based on bag-of-visual-words (BOVW) to mobile device are their computation cost for feature extraction and quantization and the memory storage for features and vocabularies. In this paper a totally randomized approach based on virtual subwindows and random ferns is proposed to evaluate image's BOVW vector fast and with low memory consumption;cross entropy is used to measure the similarity' between BOVW vectors which are evaluated by m-estimate method to improve classification accuracy;inverted file is involved to reduce the memory usage of training results and accelerate calculation in query stage. Experiments show that this method is simpler, faster, more compact than previous methods while maintain high classification accuracy.