A biologically-inspired concept for active image recognition

A novel concept is proposed that uses active information gathering for recognizing objects on images. This novel concept mimics recent neurobiological insights on human control of eye movements (TD algorithm). TD algorithms are predictive reinforcement algorithms for learning of action sequences. In the proposed framework, standard techniques, such as template matching, are used as processing steps. Each processing step compares a template of one part of the object with image locations and computes a value that describes how well the template matches. A TD algorithm is trained on many images to optimize the sequence of processing steps by providing feedback whether the final object recognition was correct or incorrect. After training, the algorithm searches for template matches with the sequence of templates and locations that are most promising for recognition of a certain object. This object recognition strategy resembles active information gathering by saccadic eye movements.