Hierarchical Feature Extraction with Sparseness Constraint

We proposes a hierarchical feature extraction method, the sparse neural response, motivated by the neuroscience of the visual cortex. The proposed method builds an increasingly complex image representation by alternating between a sparse coding and a maximum pooling operation. Generally speaking, each sample (image patch or low layer sparse neural response) and its neighbors lie on or close to a locally linear patch of the manifold. For a given sample, the desired coding is sparse, since it is only represented in terms of its neighbors. Consequently, sparse coding aims at obtaining new representation on the manifold. Maximum pooling is performed for achieving invariance to variations. In addition to promoting the performance of the method proposed, we also devise an effective template selection method. In this way, we can obtain discriminative features that are especially suited for recognition. Key advantages of this approach include simple implementation, fairly good discrimination ability, good robustness, and low sample complexity. A set of experiments on publicly available databases clearly indicate that the proposed approach is very promising.