Jul. 06, 2009
TopstoriesTopics

Beyond the Noise

Putting Camera Standard EMVA 1288 to Use

  • Set-up in the dark roomSet-up in the dark room
  • Set-up in the dark room
  • Signal-to-noise ratio as an independent measure for the sensitivity

Pepperl+Fuchs is a worldwide operating manufacturer of sensors and components for process and automation technology. Their portfolio encloses, among others, camera-based systems such as vision sensors and optical identification systems, which are being developed at daughter company Omnitron located in Griesheim, Germany. At Omnitron, within the framework of a diploma thesis, a procedure was introduced for the qualification of the camera sensitivity taking into account the influence of noise. Embedded in this procedure is the camera standard EMVA 1288.

EMVA 1288, hosted by the European Machine Vision Association, was originally developed to enable the users of industrial and scientific cameras to compare the technical data of the camera data sheets provided by the camera vendors. The standard provides a defined test frame as well as the defined presentation of the resulting data. Within this standard, one method is the "photon transfer method" (PTM). The work for the diploma thesis employed the EMVA 1288 standard to ensure a later comparability of the results.

The Formation of Noise

As an introduction to the standard, this section will briefly explain the functionality of a camera. Light, in the form of photons, shines on the image sensor. The photon transfer creates an electrical charge in the image sensor. With the help of a capacitor, this charge is converted into a voltage, which is amplified before an ADC digitizes it. The digitized steps can be shown as grey values in digital images.

In this process a noise will develop that adds to the signal. This noise is comprised of in three main types: a temporal noise, a spatial noise and quantization noise.

As to the temporal noise, the digital signal of an individual pixel varies from frame to frame. Reasons for the spatial noise are due to the production process of the image sensor. In an image one can recognize it as a fixed pattern. It also exists when every pixel is exposed to precisely the same amount of light and the temporal noise is removed. Particularly CMOS image sensors suffer strongly from spatial noise. Quantization noise stems from converting the amplified voltage into a digital signal with discreet steps: The coarser these steps are the bigger is this specific noise.

Standardized Measurement Set-ups

The standard EMVA 1288 describes a complete set-up for the experiment and the measuring requirements.

First, it is important to shield the camera against any ambient light and to work with a defined monochrome illumination. The wavelength can be chosen arbitrarily. Moreover, the image sensor should be illuminated very homogeneously. This is accomplished best by a set-up without an optical lens and by making use of a Lambert's emitter. The standard´s description of the geometrical configuration is comprehensive and very easy to understand. The diameter of the emitter in combination with the distance to the camera should result in an f-number of 8. Figure 1 shows the set-up used in our experiment.

The light source is an approximate Lambert's emitter with a diameter of 25 mm. Consequently, the distance to the camera is 200 mm. It is important that the geometrical centers of both components are positioned along the same axis and that they stay parallel to each other.

In addition the following settings on the camera are necessary:

  • The grey level resolution (number of bits per pixel) must be as high as possible. Through a higher bit number the ADC receives more steps to discretize the analogue signal. The higher the resolution, the lower the quantization noise.
  • The amplification must be as low as possible but high enough to ensure that the noise is greater than or equal to one grey value.
  • The Offset (Black Level Calibration) must be as low as possible but high enough to ensure that the signal in darkness is greater than or equal to one grey value.
  • No automatic corrections and/or settings may be used.
  • The measurement has to be executed in the linear mode of the camera.

Signal and Noise

According to standard EMVA 1288, the signal of an image is represented by the average of all grey values of the image.  

Two noise sources dominate the amplitude of the signal, the "Photon Noise" and "Amplifier Noise". Photon Noise originates from the nature of the light itself. The photons do not hit regularly onto the image sensor. The irregularity is given by a certain mathematical expectation. A Poisson distribution determines this expectation.

The Amplifier Noise is a white noise and is added when the analogue signal is amplified.

Both dimensions belong to the type of temporal noise. Therefore, in order to scrutinize the sensitivity, only the temporal noise needs to be observed.  

The remaining inhomogeneous conditions, the FPN (Fixed Pattern Noise) and PRNU (Photo Response Non Uniformity) are removed by taking the difference from two images. FPN originates from the variation of the dark current from pixel to pixel. PRNU originates from a different photosensitivity of the individual pixels. A local pattern develops, which is why these noises belong to type of spatial noise.

Calculation of Sensitivity

The signal and the noise are the basic factors to determine all data sheet values depending on the temporal noise. The sensitivity is one of the data sheet values that can be calculated. More data sheet values with the calculation formula documented by standard EMVA 1288 are given in table 1.

The sensitivity shows how many electrons are needed to generate one digital unit. It is calculated with the help of the reciprocal value of the "Total Conversion Gain K" (TCG). The TCG describes a conversion factor between the electrons generated in the image sensor and the digital signal.

Image sensors are dependent on temperature. This means that also a thermal effect releases electrical charges in the image sensor. But here, only the charge caused by light is of any interest. Therefore the signal and the noise are corrected with values from respective dark images.

However, sensitivity is not an independent measurement, but influenced by amplification of voltage. Although this ought to be set as low as possible, it can, however, vary even when using the same image sensors in different cameras. Hence, in this case the amplification should be declared. This is, though, rather contra-productive in the course of the required comparability of data sheets. Comparing the defined signal-to-noise-ratio (SNR) is a sensible alternative.

Because the noise is enhanced as strongly as the signal, it is eliminated by creating the quotient. Thus, it can be determined how many photons are necessary in order to achieve a defined illumination.

Test Procedure

The images for the evaluation stem from two series. First a series with a defined illumination is taken at varying exposure times. Variation of exposure times must include from area SNR = 1 up to the saturation area of the camera. Also, it is necessary to provide two images (see signal and noise) from every exposure time.

Once the illuminated series is finished, it is most feasible to switch off the lighting and start the second series with the same exposure times (see calculation of the sensitivity).

The signal and the temporal noise are calculated for all images.

Measuring Results for Sensitivity

For graphical representation of the TCG, the corrected temporal noise is plotted versus the corrected signal. The gradient of the best-fit line through all points represents the TCG. In this example it is K = 0.099 and the matching reciprocal value D = 10.1e-/DN.

Figure 2 shows the comparison of the SNR. The ratio of the comparison is 40:1. According to ISO 12232, this can be considered as an excellent illumination. In a graph along logarithm dualis this would be an SNR of 5.3 bits. To reach this SNR, approx. 2,900 photons are necessary.

An overview of the results dependent on temporal noise is shown in table 1. These results are based on the wavelength 617 nm, which was chosen, because it is the standard wavelength of the integrated LED light of the examined camera hardware.

 

 

Quantum efficiency

46.5

%

Sensitivity D

10.1

e-/DN

Saturation μp,sat

20,500

photons

Total Conversion Gain K

0.099

DN/e-

                                                           Table 1: Overview of the results

 

 

Table 2 shows the derived parameters.

 

 

SNR

46.5

%

DYN

9.86

Bit

Absolute sensitivity threshold μp,min

22

photons

Table 2: Derived results 

It can be concluded that introducing a standardized data sheet for cameras is an important step. Not only is it an advantage for users to choose the right camera for an application, but it also offers a benchmark possibility for suppliers developing new products.

Pepperl+Fuchs is able to compare different camera configurations in a fast and efficient way with the described measurement set-up. During the work for the diploma thesis the standard EMVA1288 could be easily implemented, with very little investments and no problems. The quoted literature in the standard was of great help.

 

All formulas of this text can be viewed in the pdf file.

Authors

Contact

Pepperl+Fuchs GmbH
Lilienthalstrasse 200
68307 Mannheim
Germany
Phone: +49 621 776 0
Telefax: +49 621 776 1111

Register now!

The latest information directly via newsletter.

To prevent automated spam submissions leave this field empty.