Towards an evolved color-light enhancement algorithm for low-light skin segmentation in a highly unstable lighting environment

Abstract

Detecting skin color and lights seems to be a very easy task that the brain can do with a very high level of accuracy and in the fundamental level of the vision system. In the opposite side, the machine is struggling to describe light and skin color and to find a compromise between the best color space reprentation and the optimum light enhancement method. So dealing with light have to be taken very seriously in the start of any vision system, since the most of confusion or misunderstanding are caused by ignoring light as the mean feature in color, shapes, textures and almost every object. Without light, we can’t see, and incorrect descrip- tion and/or treatment can lead to exponential degradation of results in detection, tracking, and recognition of objects. Several applications, such as face processing, computer-human in- teraction, human crowd surveillance, biometric, video surveillance, artificial intelligence and content-based image retrieval etc. All of these applications stated above, require skin detec- tion, which is often simply considered as a preprocessing step, for obtaining the "object/face". In other words, many of the techniques are proposed for these applications assume that, the location of the skin is a simple task and pre-identified and available for the next step, but in the opposite skin tones are not easy to find especially in some complex conditions, one of the hardest problems to surpass is low-light Illumination because it’s an important factor in determining the quality of images and also can have much effect on the evaluation. A wide variety of skin segmentatin methods have been proposed in recent years. However, most of them they don’t take the low light problem seriously. And they assume that skin patches are readily available for treatment. We do not receive images with just skin or faces. We need a system, which can detect, locate and isolate them, so thzy can be given as input to deal with recognition systems, in real contexts with complex lighting environments and for a big variety of ethnic skin differences. we propose a new method for skin detection in a complex lighting environment, based on hybridaton of: first, Improved retinex light enhancement “Im- proved MultiSclae Retinex with Color Restoration (MSRCR-SCB), so the first is based on light compensation and the second is based on color amelioration”, and second, multi-skin region detector based on the most famous color spaces (Ycbcr, RGB, HSV), and the choice is based on the robustness if each one in some situation. This enhancement or compensation of light consumes little computing time and conducts us to surpass a huge problem of chrominance and edges or even shapes that are deeply affected by the low light degradation, and the result is images to be used as an input in the recognition process, in order to improve the performance. This approach will be implemented and tested on a number of challenging low-light public databases (pratheepan/SFA/dark FACE/humanae/MASKcelebA. . . ), and compared over sev- 3 eral state-of-the-arts in terms of enhancement quality and efficiency and restore the natural colors into the images taken in real-world situations or darker lighting conditions.