Round Trip Imaging

Introduction

With the prevalence of cameras on smartphones, the quality of every-day-use cameras has improved immensely. What was once available only through an expensive camera-only device is now available in your pocket. This prompted me to ask the question, how good does a camera have to be? Sure, the higher the quality the better the result, right? Well let us answer that question with an ISET experiment!

Background

The goal in this computational experiment was to identify how the quality of a camera affects the resulting (output) image of a predetermined image source. The source and output display remained constant while only the camera sensor parameters were varied. The quality of the camera in this study is defined by the camera read noise and the resolution of the camera sensor, specifically pixel height and width.The predetermined image source is an image that was captured using an iPhone 6 camera (1/3-inch sensor, 8 MP) shown below. The image contains typical text that a viewer might read. Since the readability of text is subjective, this experiment only yielded a qualitative measure of readability. It did however provide quantitative results on the readNoise and pixel resolution of real camera sensor modules.

Methods

The camera simulation was completed using standard functions in the ISET toolkit. The figure below diagrams the imaging pipeline. First an image is taken. That image is then converted into a scene. Various cameras are then created by modifying readNoise levels and pixel resolution in the camera.sensor structure. Each one of those cameras then captures an image and the output is displayed for the viewer.

Results

readNoise Effects

At first a wide range of noise levels was studied. Although, it was difficult to assess noise levels without any real camera data (default readNoise level in ISET is 1e-3). I searched camera module literature to identify down some of the real noise levels in operational camera modules. I was able to find two cameras with comparable signal-to-noise (SNR) and dynamic range (DR) levels. These cameras are typical cameras used for phones and rear-facing laptop cameras. The table below contains SNR and DR data for the two cameras.[1-2] It was difficult to translate readNoise levels in Matlab to dB levels in SNR and DR in the manufacturer's literature. I estimate that 70 dB SNR is approximately 1.5e-3 in readNoise. This is due to the fact that the SNR ratio for a 70 dB SNR is approximately 1/(1.5e-3).

Table 1 - Camera Data

Camera	SNR	DR	Cost ($)	Resolution
Aptina	37	70.8	5.08	720p
e-con	36	68.0	19.0	1080p

For the simulations, I varied readNoise near the 1.5e-3 levels. Although the text is visible in all three of the images shown below, the image with the highest readNoise level (3e-3) is not quite readable, at least not entirely. My conclusion here was that a readable amount of noise must fall somewhere below readNoise = 3e-3. This coincided with my previous estimate and therefore a redNoise level of 1.5e-3 was used for the following simulations.

Pixel Resolution

Working directly with the sensor structure in Matlab allowed me to understand some of the subtleties in sensor-to-display data. Instead of working with the rows and columns of the sensor, I instead focused on the pixel size to achieve improved resolution. Given a readNoise level of 1.5e-3, I now wanted to see if improved sensor resolution was in fact worth the nearly 400% increase in cost of the e-con camera module. The pixel width and height for each camera were equal. The pixel widths for the low and high resolution simulations were 2.8 and 1.9 micron, respectively. The images below are the resulting camera outputs using an sRGB display.

References

[1] Aptina Imaging Data Sheet, "MT9M114: 1/6-inch 720p High Definition (HD) System-On-A-Chip (SOC) Digital Image Sensor," http://www.datasheetlib.com/datasheet/1269108/mt9m114_aptina-imaging-corporation.html

[2] e-CAM52A_5640_MOD - 5 Mega Pixel Camera Module, https://www.e-consystems.com/5MP-parallel-camera-module.asp