Visual attention engine for low energy visual object recognition

Abstract

A digital Cellular Neural Network block named the Visual Attention Engine that executes visual attention algorithms to enable low energy object recognition is presented. In the first half of this thesis, an attention driven approach to view-based visual object recognition for low-energy real-time applications is presented. In this approach, the steps of local feature extraction and matching in existing object recognition algorithms are augmented by an attentional mechanism that is based on Cellular Neural Networks. During the local feature extraction stage, the mechanism assigns the limited resources of a vision system to the most meaningful regions of an image and regulates the amount of effort exerted to each region based on a saliency map. During feature matching and subsequent object matching, matching speed is optimized by intelligently controlling the accuracy of database queries. The design and implementation of the Visual Attention Engine are presented in the second part. It accelerates saliency map generation, which is a computationally intensive process requiring processing at the global scale. The Visual Attention Engine is an $80 \times 60$ Cellular Neural Network based on a two-dimensional shift-register architecture that nearly eliminates data access overhead. 120 processing elements shared by the cell array provide a peak throughput of 24GOPS. The $4.5mm^2$ semi-custom block is fabricated in a 0.13um 8 metal layer process. Object recognition performance analysis on the COIL-100 dataset shows that total energy consumption is reduced by over 70% with only a negligible impact on matching accuracy by employing the VAE.