CMOS Vision Sensors: Embedding Computer Vision at Imaging Front-Ends

Abstract

CMOS Image Sensors (CIS) are key for imaging technol-ogies. These chips are conceived for capturing opticalscenes focused on their surface, and for delivering elec-trical images, commonly in digital format. CISs may incor-porate intelligence; however, their smartness basicallyconcerns calibration, error correction and other similartasks. The term CVISs (CMOS VIsion Sensors) definesother class of sensor front-ends which are aimed at per-forming vision tasks right at the focal plane. They havebeen running under names such as computational imagesensors, vision sensors and silicon retinas, among others. CVIS and CISs are similar regarding physical imple-mentation. However, while inputs of both CIS and CVISare images captured by photo-sensors placed at thefocal-plane, CVISs primary outputs may not be imagesbut either image features or even decisions based on thespatial-temporal analysis of the scenes. We may hencestate that CVISs are more “intelligent” than CISs as theyfocus on information instead of on raw data. Actually,CVIS architectures capable of extracting and interpretingthe information contained in images, and prompting reac-tion commands thereof, have been explored for years inacademia, and industrial applications are recently ramp-ing up.One of the challenges of CVISs architects is incorporat-ing computer vision concepts into the design flow. Theendeavor is ambitious because imaging and computervision communities are rather disjoint groups talking dif-ferent languages. The Cellular Nonlinear Network Univer-sal Machine (CNNUM) paradigm, proposed by Profs.Chua and Roska, defined an adequate framework forsuch conciliation as it is particularly well suited for hard-ware-software co-design [1]-[4]. This paper overviewsCVISs chips that were conceived and prototyped at IMSEVision Lab over the past twenty years. Some of them fitthe CNNUM paradigm while others are tangential to it. Allthem employ per-pixel mixed-signal processing circuitryto achieve sensor-processing concurrency in the quest offast operation with reduced energy budget.