Original | Processed

Many consumer cameras, especially cameras with small image sensors or pixels (mobile imaging devices and point-and-shoots), have signal processing that varies over the image plane. Sharpening is applied near contrasty features (like edges), but noise reduction (lowpass filtering) is applied— often strongly— the absence of sharp features, resulting in loss of texture detail. Such cameras will perform well on slanted-edge tests while producing unsatisfactory images. To emphasize this we show a real camera phone image, where the window and shingles have been strongly sharpened, but texture in the pine shrubs has been completely lost.Imatest measures texture sharpness in two modules:

  • Log F-Contrast, which relates pattern contrast to texture loss, and
  • Random/Dead Leaves, which supports two types of chart. The Scale-invariant random pattern minimizes sharpening and maximizes noise reduction. The Dead Leaves pattern is more representative of typical images. The Dead Leaves pattern has attracted considerable attention from the industry, particularly from the Camera Phone Image Quality (CPIQ) group. Imatest’s new Spilled Coins (Dead Leaves) chart has several strong advantages over existing charts.

Texture Examples

Images used in Random / Dead Leaves

this section illustrates images analyzed in Random Scale-invariant & Dead Leaves. The images are not shown original size; they’ve been resized to be approximately equal in magnification with respect to the original chart image— with enough magnification to show the results of the camera optics and image processing.

Original pattern (cropped from the middle of the chart), for reference. Original pattern (reduced from file used to print the chart)
Panasonic DMC-G3 in-camera raw, converted by dcraw with default quality setting.This image has no sharpening or noise reduction. Panasonic DMC-G3 converted from raw image by dcraw
Panasonic DMC-G3 in-camera JPEG, default settings
Typical of in-camera JPEGs, some sharpening is visible. Noise reduction is not obvious.
Panasonic DMC-G3 JPEG
Early 8 megapixel camera phoneThis image has some serious issues: poor sharpness, significant texture loss in lowest contrast areas, but it also shows sharpening-related artifacts. The left side of the image (not visible here) suffers from a serious sharpness loss.
Recent 8 megapixel camera phones (2012) have much better performance.There are some significant issues with the Spilled Coins/Dead Leaves analysis for this camera, shown in detail below. Because there seem to be a lot of artifacts around edges in the image, the Spilled Coins MTF is not bad, even though fine texture is quite absent from the image. 8 Megapixel camera phone
Camera phone MTF from Spilled Coins pattern (looks pretty good, but…)Here is the Spilled Coins MTF from the 8MP Camera phone. Results are pretty good, but the high frequency energy is from artifacts around edges: it is not representative of true texture detail. It would seem from this image that the Random Scale-Invariant chart might produce results that are more representative of the texture blur.
Camera phone MTF from Spilled Coins pattern
Camera phone MTF from slanted-edge. Flunnks badlyHere are the results from the slanted-edge at the bottom of the image. Results are quite poor, as expected, with little meaningful energy above 0.3 C/P. It’s a good bet that this ill-behaved camera was not included in the tests that established strong correlation from Dead Leaves-derived MTF and perceptual image quality.
Camera phone MTF from slanted-edge.

Spilled Coins vs. Slanted edge MTF

This section demonstrates how demosaicing is the cause of a commonly observed discrepancy between Spilled Coins and slanted-edge MTF measurements.

When MTF results are compared for the Spilled Coins pattern and slanted edges from the same image, we often see a discrepancy at high spatial frequencies— typically above 0.3 or 0.4 cycles/pixel, depending on the image. The Spilled Coins MTF has a floor, i.e., it doesn’t drop below a minimum value as spatial frequency increases, while the slanted-edge MTF continues to drop into the noise. We have found that this result is primarily caused by demosaicing, which involves nonlinear processing that treats edges (in the Spilled Coins pattern) differently from noise (in the smooth areas on the sides of the Spilled Coins pattern).

Summary: We have found that discrepancies between Spilled Coins and slanted-edge MTF measurements are primarily caused by demosaicing, which is a type of nonlinear processing that handles edges and random noise quite differently.

Spilled Coins MTF Slanted-edge MTF
Demosaiced First, we analyze a demosaiced image taken at f/22 that doesn’t have much high frequency energy. this makes it useful for looking at a single color plane of an undemosaiced image (shown below).

Spilled Coins MTF, demosaiced, f/22, ISO 160

Agreement between the two methods is excellent below 0.25 C/P, but they diverge at higher spatial frequencies. Similar results are observed at ISO 800, but curves are rougher.


Slanted-edge MTF, demosaiced, f/22, ISO 160
No demosaicing Next, we analyze a single color plane of the same image, undemosaiced. The number of vertical and horizontal pixels are half that of the standard demosaiced image, and the pixel pitch is double. Hence the frequencies are approximately doubled for this curve, e.g., MTF @ 0.4 C/P undemosaiced ̃ MTF @ 0.2 C/P demosaiced.

Spilled Coins MTF, demosaiced, f/22, ISO 160

Agreement between the two methods is excellent.

Slanted-edge MTF, undemosaiced, f/22, ISO 160
No demosaicing Finally, we analyze an image that has not been passed through a camera. It is an original image used to create the Spilled coins chart that has been reduced, linearly sharpened (USM R=2), and linearly blurred (gaussian R=1). By linear we mean perfectly uniform processing with no thresholds.

Spilled Coins MTF, original image reduced, USM sharpened (R=2), blurred (gaussian R=1)

Agreement between the Spilled Coins and slanted-edge pattern is nearly perfect.

Slanted-edge MTF, original image reduced, USM sharpened (R=2), blurred (gaussian R=1)