Integrated Cmos Image Sensors For Computational Imaging

Abstract

Low cost cameras have proliferated in the last decade by taking advantage of the semiconductor technology (CMOS) used for the manufacture of computer processors and other integrated circuits. An image sensor is the key integrated circuit component in a camera that serves the role of faithfully recording the interaction of light with the world. Fabrication in a CMOS process allows a variety of passive and active components to be included with the image sensor. This manufacturing flexibility combined with increasing availability of computational power opens new possibilities for image sensing. This dissertation presents two CMOS image sensors intended not for direct image capture but to enable information processing tasks such as post-capture refocusing, depth mapping, feature extraction, optical flow, and velocimetry. A major portion of this thesis advances a recently introduced light sensing device, called angle sensitive pixel (ASP), which selectively responds to the local incident angle of light rays. Although the previously presented device enabled post-capture refocus and range finding by capturing light fields it was limited by its poor optical sensitivity due to its use of metallic gratings. In this work, I report on a new set of structures for angle detection with significantly better optical efficiency. For the fabrication of these devices, I develop a post-CMOS process flow for etching phase gratings at each pixel site and demonstrate its scalability on a 2.6 mm x 2 mm sensor. Further, the merits of an image sensor whose pixels possess carefully tailored angular responses is demonstrated by building an array of angle sensitive pixels to compute Gabor transforms in the optical domain. Images are captured by the sensor in the form of transform coefficients that readily lend themselves to data compression. An amplifier and analog-to-digital converter that exploit the sparsity of the image transform coefficients are included on the chip. I show how measurements captured by the sensor can be used to recover either a low-resolution 4D light field using fast, linear processing, or a high-resolution light field using sparsity-constrained optimization. In the latter part of this dissertation I present a CMOS image sensor for efficient computation of rotation parameters such as angular position and velocity. Unlike a conventional image sensors built using a rectangular grid of pixels, the array introduced here uses circular photodiodes arranged uniformly in polar coordinates to enable rapid image based measurements on revolving targets.