A Multispectral Vision Sensor with Embedded Convolutional Neural Network Using Programmable Fractional Weights and nMOS-Only PWM Pixels

This paper presents a multispectral vision sensor integrating convolutional neural network (CNN) for color-based vision tasks. The sensor utilizes fractional convolution (Conv) weights with a resolution of 0.1 using weighted exposure control for precise adjustments. It features 5b fully connected (FC) weights enabled by an analog multiply-and-accumulate (MAC) unit, optimizing both area and energy efficiency. The proposed nMOS-only PWM pixel improves the fill factor by 30.6% compared to prior works. A <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$258 \times 129$</tex> CMOS 180 nm sensor array consumes <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$137.5 \mu ~\mathrm{W}$</tex> at 20 fps per channel with iFoM of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$206.6 \text{pJ} /$</tex> pixel <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\cdot$</tex> frame, achieving 90.9% and 89.6% accuracy on diverse detection tasks.