Variable layout CMOS pixels for end-to-end learning in task specific Image Sensors

CMOS Image Sensors produce data used in several machine learning models. Typically, their layout is limited to rectangular grid of uniform pixels. However, their shape and size can provide another optimisation parameter in end-to-end training in specific machine learning tasks. Variable layout pixels are observed as one potential approach. An image sensor with such variable and learned layout is presented. A technique for automatic generation of such layouts is also described, along with data from a typical implementation of such pixels.