Multi-class Image Classification Based on Fast Stochastic Gradient Boosting

Li Lin, Wu Yue, Ye Mao


Nowadays, image classification is one of the hottest and most difficult research domains. It involves two
aspects of problem. One is image feature representation and coding, the other is the usage of classifier. For
better accuracy and running efficiency of high dimension characteristics circumstance in image classification,
this paper proposes a novel framework for multi-class image classification based on fast stochastic
gradient boosting. We produce the image feature representation by extracting PHOW descriptor of image,
then map the descriptor though additive kernels, finally classify image though fast stochastic gradient
boosting. In order to further boost the running efficiency, We propose method of local parallelism and an
error control mechanism for simplifying the iterating process. Experiments are tested on two data sets:
Optdigits, 15-Scenes. The experiments compare decision tree, random forest, extremely random trees,
stochastic gradient boosting and its fast versions. The experiment justifies that (1) stochastic gradient
boosting and its extensions are apparent superior to other algorithms on overall accuracy; (2) our fast
stochastic gradient boosting algorithm greatly saves time while keeping high overall accuracy.

Full Text:


Creative Commons License
This work is licensed under a Creative Commons Attribution 3.0 License.