I am doing machine learning for Object Detection, my data set would never cove all day to consider the effect of illumination
Taking only the gray level into consideration, a first approach is to consider the change in the average pixel value of the image when the scene ilumination changes. The ability of the hardware to compensate for this varies enormously among different models of camera, lens and illumination. If the hardware does not compensate for this, the average gray level of the image will migrate to one of the ends, 0 or 255, depending on the lack or excess of illumination. In both cases, the image will be more noisy (random values added to gray level) and the contrast will diminish (smaller standard deviation of the pixels values). Both conditions result in loss of detail in the image.
A second very important factor to take into consideration is the ability of the camera to properly encode scenes with high contrast. As an example the left side of the image could be showing the ground iluminated by the Sun at the same time the right side of the image is presenting the shadow of a wall. Only very good hardware will be able to compensate this. If the compensation is poor, the image will end showing all-white on the left or all-black on the right, making it very hard to distinguish objects.In terms of the gray level values, this codition will diminish the contrast (standard deviation) of part of the image.
These effects will change with the position of the light source (ex.: Sun along the day), or when the light source is point like (sunny day) or is spread (as in a cloudy day).
As an image is generated on the plane as multiplication of amount of illumination and amount of light reflected back from an object, when the illumination is very low, the intensities of the pixels will result in to the lower values.
illumination it is a problem for computer vision and it introduces also shadows. If you want to avoid those problems, you can change the space color to HSV or YCbCr and exclude the channel of the illumination.
Other method is to obtain a local descriptor to those images, for instance, LBP, HOG...
Surface reflectance (as captured by a camera) is a characteristic light + object properties. Sunlight is characterized by a standard value of CRI(color rendering index) . This parameter tell us how well has color been captured .
For example: A light source (white light)with spectral characteristics or wavelength more towards yellow, will make black look like brown.
It is also dependent on temperature of source (around 5000K for daylight)
http://www.schorsch.com/en/kbase/glossary/cct.html
That is the reason why pseudo colors vs true colors are looked into (considering illumination conditions ) .
Illumination is very important part in visible light images. The acquired image is modelled as product of reflectivity and illumination as the illumination change, the image pixel intensity will change. That's why for getting good image, the illumination is on in smart camera. Similarly high illumination is required by high speed camera. If you are using specific algorithm for your application then modify that with respect to available illumination or use controlled illumination. You can refer - https://www.researchgate.net/publication/293731256_A_Contrast_Enhancement_Technique_for_Low_Light_Images
Conference Paper A Contrast Enhancement Technique for Low Light Images
- Badges
- Science topic