In [Espejo et al. 94e, Espejo et al. 94b, Espejo et al. 92, Espejo et al. 93b, Espejo et al. 93a, Espejo et al. 94d, Espejo et al. 94c, Rodriguez-Vazquez et al. 96] Espejo et al. describe cellular neural network-based (CNN) chips designed for image processing. The main advantage of CNN for VLSI implementation is the locality of interconnections. Each cell in a CNN only connects to its nearest neighbors. The modified equation governing the behavior of a CNN is given by:
where g() is a nonlinear term defined by:
x is the state of the cell, D is an offset factor, A and B are weight factors, u is the input, and y is the output given by
Different terms in these equations can be easily implemented using current mode circuits. The dx/dt term has been implemented using a current mode integrator in the feedback loop.
Two special chips with different connections and weights have been
designed and fabricated in a 1.6 
m 1P-2M CMOS process. The first
one contains a 16 
16 array for detecting connected components
(DCC). The chip dimensions are 2.5mm 2.5mm, and dissipates 42
mW. Each cell in this chip occupies 
. The
second chip performs Radon transform on a 16 16 image. In this
chip the input can be selected from external or internal (optical)
sources. The chip operates in sampled-data mode. The chip area is
2.67mm 2.68mm, and its power dissipation is 330 mW. Cell
dimensions are 
.
Both chips use the multisensitivity photodetector in darlington mode (See Figure 2.23). The circuits operate in the above-threshold region (as opposed to subthreshold) to achieve better mismatch. However, other factors such as power dissipation and transistor sizing have been traded off.
A later design with more flexibility and functionality uses a
0.8 m process with 20 22 cells in an area of
30 
[Rodriguez-Vazquez et al. 96]. This chip also uses darlington
connected phototransistors. The weights can be loaded into the array.
Therefore, the chip can be programmed to perform different functions.
It has successfully demonstrated operations such as low-pass image
filtering, corner and border extraction, hole filling, and motion
detection.