A countinuous-time cellular neural network chip for direction-selectable connected component detection with optical image acquisition

Abstract

This paper presents a continuous-time Cellular Neural Network (CNN) chip [1] for the application of Connected Component Detection (CCDet) [2]. Projection direction can be selected among four different possibilities. Every cell (or pixel) in the 32 x 32 array includes a photosensor circuitry and an automatic tuning circuitry to adapt to average environmental illumination. Electrical image uploading is possible as well. Input pixel-values are stored on local memories (one per cell), allowing sequential processing of the acquired image in different directions. The prototype has been designed and fabricated on a standard digital CMOS technology: 1.6mm, n-well, single-poly, double-metal. Circuit implementation is based on current-mode techniques and uses a systematic approach valid for any CNN application [3]. Cell dimensions, including the CNN processing circuitry, the photosensor and the adaptive circuitry are 145 x 150 mm2, of which the sensor and adaptive circuitry amounts to ~15% of the total pixel area and the wiring and multiplexing (required for direction selectability) to about 40%. The remaining 45% corresponds to the CNN processing circuitry. Pixel density is ~46 cells/mm2, and power dissipation is 0.33mW/cell. These area and power figures forecast single-die CMOS chips with 100 ´ 100 complexity and about 3W power consumption.