An Improvement of Image Recognition Performance Using Ensemble Learning for Conventional HOG Features Calculated with Different Sizes of Cell and Block

Abstract

In object recognition problem, HOG features and some machine learning methods for learning them to identify an object are known as extremely effective techniques. Recently, emphasizing feature variance and multiplying feature resolution have also been proposed for further improving recognition performance. This paper calculates multiple types of the conventional HOG by combining their various cell sizes and block sizes in an appropriate range, and proposes a simple method to learn them and evaluate the corresponding object recognition performance. Especially in evaluation experiments using real images of faces and bodies, the experimental results show that learning performance of the proposed method comparable to the extremely complicated latest methods such as CoHOG and MRCoHOG can be achieved, although its execution time slightly increases compared to the conventional HOG.