A Smart Camera Architecture with Keypoint Description and Hybrid Processor Population

Abstract:
Increasing local processing capability of smart camera systems enables better scene analysis and more accurate decisions, and this leads to expand their application areas such as wireless sensor networks. Image understanding in smart cameras has been enhanced by the incorporation of keypoint detection and description mechanisms. Small memory footprint and high matching speed of Local Binary Descriptors (LBD) make them suitable to use in embedded applications and they also have several hardware implementations, leading them to meet real-time constraints. However, the flexibility of such available hardware is quite limited. On the other hand, it is known that biological neural networks which are naturally capable of solving these problems outperform today’s electronics systems. Mimicking these structures with Spiking Neural Networks (SNN) gives very promising results, but this implementation requires special manufacturing process and lack of cost-efficiency in the current technology. Alternatively, Cellular Neural Networks (CNN) which have low-cost digital implementations, can boost the existing image analysis capabilities. In this paper, we extend our smart camera architecture with a modified CNN-based structure with 2-different types of cells showing either inhibitory or excitatory behavior. Cell connections are determined by the cell types stored on a binary identity matrix. Thus, this flexible network can be configured to incorparate with existing keypoint detection and description blocks. In addition, it also enables further analysis with complex comparisons and multiple iterations.