Large Scale Visual Classification via Learned Dictionaries and Sparse Representation

We address the large scale visual classification problem. The approach is based on sparse and redundant representations over trained dictionaries. The proposed algorithm firstly trains dictionaries using the images of every visual category, one category has one dictionary. In this paper, we choose the K-SVD algorithm to train the visual category dictionary. Given a set of training images from a category, the K-SVD algorithm seeks the dictionary that leads to the best representation for each image in this set, under strict sparsity constraints. For testing images, the traditional classification method under the large scale condition is the

k

-nearest-neighbor method. And in our method, the category result is through the reconstruction residual using different dictionaries. To get the most effective dictionaries, we explore the large scale image database from the Internet [2] and design experiments on a nearly 1.6 million tiny images on the middle semantic level defined based on WordNet. We compare the image classification performance under different image resolutions and

k

-nearest-neighbor parameters. The experimental results demonstrate that the proposed algorithm outperforms

k

-nearest-neighbor in two aspects: 1) the discriminative capability for large scale visual classification task, and 2) the average running time of classifying one image.