Analog-digital hybrid vision chips for motion processing systems based on direction-selective neural networks

Yoshihito Amemiya

We developed analog-digital hybrid CMOS circuits producing optical flows of visual images, aiming at the development of high-speed and compact motion processing systems. The circuit computes the optical flows in real time by parallel operations of a number of local motion processors on the basis of the mechanisms of biological motion detectors.

Three kinds of CMOS circuits that are useful for constructing motion-detection neurochips were developed, i.e., i) current-mode quantizers for image quantization, ii) current-mode XOR circuits for binary edge detection, and iii) analog circuits for motion detection. Figure 1 shows sample structure of the motion-detection chip consisting of the above circuits.

Fig.1(35k PNG image)Click to zoom up

The analog motion-detection circuit, which implements the unit cell of correlation neural networks, produces direction selective voltages for moving current spots between two adjacent input terminals (Fig. 2). The circuit consists of i) a delay circuit that produces delayed voltage from the input current (P1) and ii) a correlation circuit that multiplies the delayed voltage and the input current (P2). We fabricated prototype chips consisting of above basic circuits. Figure 2 shows example operations of the motion detection circuit. The output of the circuit was 0 V when the current spot moved from the terminals of P2 to that of P1, while it outputted nonzero voltage when the spot moved opposite direction, as expected. Thus, the circuit produces direction-selective responses for moving objects. These circuits can be used to construct analog-digital hybrid neuromorphic chips.