Image recognition improvement

Abstract

Bag-of-features (BoF) has already become one of the most popular models in image

classification over the years. Applying spatial pyramid matching (SPM) together with

BoF has shown to achieve much better classification accuracy and was widely used

in image categorization. In recent years, spatial pyramid matching with sparse coding

of SIFT (ScSPM) kernel evolved from SPM with sparse coding of SIFT features and

max pooling approach greatly improve the image classification accuracy basing on

common image databases testing. This report presents three new approaches as the

optimization and supplementation of ScSPM kernel, which can be applied in the

spatial matching process before max pooling in ScSPM. They are respective diagonal

segmentation (DS) approach, max pooling with Gaussian parameter (GPMP)

approach and overlapping segmentation (OS) approach. The experiments on

classification accuracy of these approaches were implemented and further analysis

was conducted to discuss the effect of each approach. Finally the result showed that

the overlapping segmentation method can one step further increase the image

classification accuracy on the basis of conventional ScSPM. Diagonal segmentation

approach working together with the quadrate segmentation approach of conventional

ScSPM also achieved better performance for some databases. However, it is just a

start of ScSPM approach improvement. There is a board space in the advancement of

existing approaches and algorithms of image recognition.