A Visual Cortex-Inspired Imaging-Sensor Architecture and Its Application in Real-Time Processing

For robots equipped with an advanced computer vision-based system, object recognition has stringent real-time requirements. When the environment becomes complicated and keeps changing, existing works (e.g., template-matching strategy and machine-learning strategy) are computationally expensive, compromising object recognition performance and even stability. In order to detect objects accurately, it is necessary to build an efficient imaging sensor architecture as the neural architecture. Inspired by the neural mechanism of primary visual cortex, this paper presents an efficient three-layer architecture and proposes an approach of constraint propagation examination to efficiently extract and process information (linear contour). Through applying this architecture in the preprocessing phase to extract lines, the running time of object detection is decreased dramatically because not only are all lines represented as very simple vectors, but also the number of lines is very limited. In terms of the second measure of improving efficiency, we apply a shape-based recognition method because it does not need any high-dimensional feature descriptor, long-term training, or time-expensive preprocessing. The final results perform well. It is proved that detection performance is good. The brain is the result of natural optimization, so we conclude that a visual cortex-inspired imaging sensor architecture can greatly improve the efficiency of information processing.