Abstract:This paper presents a novel, effective way to improve the object recognition performance of a biologically-motivated model by learning informative visual features. The original model has an obvious bottleneck when learning features. Therefore, we propose a circumspect algorithm to solve this problem. First, a novel information factor was designed to find the most informative feature for each image, and then complementary features were selected based on additional information. Finally, an intra-class clustering strategy was used to select the most typical features for each category. By integrating two other improvements, our algorithm performs better than any other system so far based on the same model.

Object Recognition by Learning Informative, Biologically Inspired Visual Features