An Object Recognition Model Using Biologically Integrative Coding with Adjustable Context

摘要

Many existing models of object recognition have a hierarchical architecture. They are based on the theory of hierarchy in brain of primate and cognitive process of human. The feature is simple in low layer while complex in high layer. However, the simple feature are local without global clues in these computational models. In this paper, we propose a novel method to code orientation feature which is local feature derived from receptive field of simple cells. The integrative coding in each simple feature, utilizing the global context information such as angle between orientations, is different from other methods of coding batch-based. This coding is scale-invariance since we overlook the distance between orientations. In addition, it is a method of feature learning since the size of context can be adjusted automatically according to special recognition task. Experimental results on ETH-80 data set demonstrate the effectiveness of our model.