Pixel Binning

From VISTA LAB WIKI

Jump to: navigation, search

Modern sensors are being designed with a variety of CFAs and circuits. Pixel binning methods compensate for the small pixel size by allowing the circuitry to combine charge (bin) across pixels. They also have the advantage of reducing the number of data points coming out of the sensor in the case that the target display is, say, a video monitor that will heavily sub-sample the output in any event.

Return to PDC Projects

Contents

[edit] Introduction

We will start simulation these methods in ISET.

see ~/SVN/ISET-4.0/scripts/sensor/binning/s_binSensorDemo.m

Implemented

  • Kodak method: vertical binning and row digital averaging
  • Decimation: blur and sample
  • Add adjacent pixel blocks
  • Average adjacent pixel blocks

Notes

  • Re Decimation: Is there an optimal "anti-aliasing" blur filter
  • What is the Phase One method?
  • Not sure what to do about Fuji EXR

[edit] Companies

[edit] Kodak binning

Kodak’s Pixelux

Brief algorithm description here


[edit] Bayer CFA binning methods implemented in ISET

Figure X. Column-wise charge binning followed by row-wise digital value averaging.
Figure X. Block-wise charge binning.

[edit] Fuji

Fuji EXR

Brief algorithm description here

[edit] Phase One

Phase One

Brief algorithm description here

[edit] News

NYT Review

[edit] Links

slides

Press article


[edit] Patents

United States Patent 6801258 – see attached CMOS integration sensor with fully differential column readout circuit for light adaptive imaging Pain, Zhou, Fossum, Cal Tech 2004 (filed 1999)

United States Patent 7091466 – see attached Apparatus and method for pixel binning in an image sensor Bock, Micron, 2006 (filed 2003)

United States Patent 7319218 – see attached Method and apparatus for pixel signal binning and interpolation in column circuits of a sensor circuit Krymski, Micron 2008 (filed 2006)

United States Patent 7402789 Methods for pixel binning in a sensor, N. Bock, Micron, 2008 http://www.freepatentsonline.com/7402789.html

United States Patent 7436442 – see attached Low light sensor signal to noise improvement Barna & Fossum, 2008 (filed 2004), Micron PATENTS


7154075 Dec 26 2006 Krymski (Micron)

Patent Number Date Inventors
6801258 Oct 5 2004 Pain, Zhou, Fossum (Cal Tech)
6878918 Apr 12 2005 Dosluoglu (Dialdg Semiconductor GmbH)
7091466 Aug 15 2006 Bock (Micron)
7319218 Jan 15 2008 Krymski (Micron)
7402789 Jul 22 2008 Bock (Micron)
7436442 Oct 14 2008 Barna, Fossum (Micron)


PATENT APPLICATIONS Patent Number Pub Date Inventors 2006/0274176 Dec 7 2006 Guidash (Eastman Kodak) 2006/0077260 Apr 13 2006 Kindt, Segerstedt (National Semiconductor) 2008/0260291 Oct 23 2008 Alarkarhu, Kalevo, Sarkijarva (Nokia) 2009/0059048 Mar 5 2009 Luo, Yang (Omnivision)*

[edit] PATENT APPLICATIONS

Kodak Patent Application – see attached Pixel binning and averaging based on illumination Filed 2006

another Kodak Patent Application CMOS image sensor pixel with selectable binning http://www.wipo.int/pctdb/en/wo.jsp?IA=US2006020586&wo=2006130518&DISPLAY=STATUS

[edit] SPIE PAPER

Comparing hybrid and digital image resizing methods on perceived image quality

Joyce Farrell, Mike Okincha, Manu Parmar, Brian Wandell

Introduction

High-resolution digital cameras can generate lower-resolution video streams by pixel-binning. This is a general term that describes the process of combining data from nearby pixels of the same color. The term can be applied to describe signal combination either in the pixel, in the analog signal chain, or in the digital domain after the ADC, alone or combination. Pixel binning reduces the number of output samples, and depending on the implementation the process can increase the signal or reduce noise. Pixel binning improves the signal-to-noise ratio (SNR) at the expense of spatial resolution.

We compare several methods of pixel-binning. For example, one method combines some pixel values on the sensor and some after digitization (hybrid). Another method combines all pixel values after read-out from the sensor array and ADC conversion (digital). We perform a set of quantitative simulations that analyze the impact of hybrid and digital down-sampling methods upon perceived image quality.

We analyze a hybrid pixel binning method described by Eastman Kodak [1] that is implemented in a 4T-pixel architecture. In this method, the charge is summed from two nearby pixels in the same color channel and column prior to readout. This step reduces the column spatial resolution of the output image so that only half as many pixels values are read-out. After quantization, the digital values from nearby pixels in the same color channel and row are averaged. This digital averaging reduces the row resolution to match the column resolution.

A purely digital implementation of pixel binning operates on pixel data after readout and quantization. The most common digital approach is to blur and sub-sample (decimate) the digital data. The blurring filter is chosen to minimize the sampling artifacts caused by down sampling. This method requires that every pixel be read out.

At low signal levels read noise is a significant factor in pixel SNR. The hybrid approach to pixel binning increases SNR by both increasing the signal and reducing the amount of the read noise in the data: there is one read for every two pixels. The digital approach increases SNR by signal averaging, but it does not reduce the read noise contribution. The digital averaging improves SNR, but the data contain twice as much read-noise as the hybrid method.

We determine whether the differences between these two methods are likely to lead to significant differences in image quality, and if so under which imaging conditions. We use both an analytic and perceptual approach. In the analytic approach we use the ISET digital camera simulator to compare the effects of these two pixel binning methods. We analyze the effects of these summing and averaging procedures in terms of SNR, color aliasing and spatial acuity.

In the perceptual approach, we determine the contributions that photon and read noise have on perceived image quality by presenting images generated with different mean scene luminance, sensor read noise and resizing methods. Images are presented in pairs and subjects are asked to indicate which of the two images they prefer. We use the number of times each image is preferred over all other images as a measure of supra-threshold image quality. The preference scores quantify the effects that uncorrelated image noise, color aliasing artifacts and spatial blur have upon perceived image quality.

ISET SIMULATIONS

  1. Vary scene luminance and read noise
  2. Use of different targets for metrics (see below)
  3. Use two faces for preference experiments

ANALYSIS

  1. SNR - how to measure it?
  2. Color aliasing - how to measure this?
  3. spatial acuity- do not like the ISO 12233 methods

IMAGES FOR PREFERENCE STUDIES

Personal tools