Reinforcement learning of informative attention patterns for object recognition

Attention is a highly important phenomenon emerging in infant development [1]. In human perception, sequential visual sampling about the environment is mandatory for object recognition purposes. Sequential attention is viewed in the framework of a saccadic decision process that aims at minimizing the uncertainty about the semantic interpretation for object or scene recognition. Methodologically, this work provides a framework for learning sequential attention in real-world visual object recognition, using an architecture of three processing stages. The first stage rejects irrelevant local descriptors providing candidates for foci of interest (FOI). The second stage investigates the information in the FOI using a codebook matcher. The third stage integrates local information via shifts of attention to characterize object discrimination. A Q-learner adapts then from explorative search on the FOI sequences. The methodology is successfully evaluated on representative indoors and outdoors imagery, demonstrating the significant impact of the learning procedures on recognition accuracy and processing time.