Large Scale Image Classification with Many Classes, Multi-features and Very High-Dimensional Signatures

The usual frameworks for image classification involve three steps: extracting features, building codebook and encoding features, and training the classifiers with a standard classification algorithm. However, the task complexity becomes very large when performing on a large dataset ImageNet [1] containing more than 14M images and 21K classes. The complexity is about the time needed to perform each task and the memory. In this paper, we propose an efficient framework for large scale image classification. We extend LIBLINEAR developed by Rong-En Fan [2] in two ways: (1) The first one is to build the balanced bagging classifiers with under-sampling strategy. Our algorithm avoids training on full data, and the training process rapidly converges to the solution, (2) The second one is to parallelize the training process of all classifiers with a multi-core computer. The evaluation on the 100 largest classes of ImageNet shows that our approach is 10 times faster than the original LIBLINEAR, 157 times faster than our parallel version of LIBSVM and 690 times faster than OCAS [3]. Furthermore, a lot of information is lost in quantization step and the obtained bag-of-words is not enough discriminative power for classification. Therefore, we propose a novel approach using several local descriptors simultaneously.