Convolutional Sparse Feature Descriptor for Object Recognition in CIFAR-10

In this work we address the problem of feature extraction for image object recognition. We propose a new, learned, feature descriptor for images, the convolutional sparse descriptor, which is based on recent advances in machine learning. It computes a spatial representation of the entire input image based on feature responses of local descriptors. The feature responses are calculated using a learned dictionary, which is learned using the sparse coding algorithm, instead of the vector quantization (VQ). Experiments on the benchmark CIFAR-10 show that our method outperforms several state-of-the-art algorithms.