BMAX: A Bag of Features Based Method for Image Classification

This work presents an image classification method based on bag of features, that needs less local features extracted for create a representative description of the image. The feature vector creation process of our approach is inspired in the cortex-like mechanisms used in "Hierarchical Model and X" proposed by Riesenhuber & Poggio. Bag of Max Features - BMAX works with the distance from each visual word to its nearest feature found in the image, instead of occurrence frequency of each word. The motivation to reduce the amount of features used is to obtain a better relation between recognition rate and computational cost. We perform tests in three public images databases generally used as benchmark, and varying the quantity of features extracted. The proposed method can spend up to 60 times less local features than the standard bag of features, with estimate loss around 5% considering recognition rate, that represents up to 17 times reduction in the running time.