AI-capable Computational CMOS Image Sensors: from Concept to Trend

The field of CMOS image sensors is rapidly evolving, driven by advancements in image processing tasks enabled by Convolutional Neural Networks (CNNs). This progress has led to the development of a specialized class of CMOS image sensors, known as Computational CMOS Image Sensors (C²IS), which are specifically designed for CNN acceleration and inference, among other applications. This paper provides an overview of these specialized image sensors, tracing their conceptual development and examining various reported architectures, highlighting their respective advantages and limitations. A central design trade-off between processing flexibility and energy efficiency is discussed, as it shapes the trajectory of innovation in this field. For each architecture, notable works are summarized, emphasizing key characteristics, limitations, and potential drawbacks. While designs are often application-driven, a comparative analysis of works across different architectures is also presented to provide a broader perspective.