Fast IR Target Shape Classifier on Focal Plane Array

摘要

We propose an original VLSI architecture to be implemented on a single chip focal plane array, dedicated to the fast recognition and classification of IR targets in unfriendly and noisy environment, using a multi-level shape extraction technique. The original algorithm described in this paper has the opportunity to use the full hardware acceleration of focal plane array for the multi-tracking mechanism. The system derives full benefit from the deep cooperation between a pixel processor matrix and a small neural network classifier implemented on the same chip. The pixel matrix can provide parallel extraction of several target shapes using any user-defined or automatic gray threshold, while the neural classifier identifies and tracks each detected target. The identification procedure can support any rotational angles and moderate zoom factors because the main information is the shape of target, not its position nor its size. Nevertheless, these last features are used to distinguish between twin targets in the same scene. Using a standard 0.6μm CMOS technology, a 1cm~2 circuit holding a 128x128 matrix can provide at least 500 multi-target classifications per second.