Shadow Removal From A Single Real Scene Image

Though digital image processing becomes more and more perfect after which has been developing for several decades, some problems such as shadow removal which refer to the mechanism of processing and understanding of visual signal conducted by human brain cannot be solved effectively. Shadow appears in color image has been exceeded the deterioration of image quality caused by traditional noise. Many image enhancement and image restoration techniques can not satisfy the requirements of shadow removal from color images. Removing shadow from a color image and supplying high-quality image for other image processing algorithms has been a research hotspot in computer vision field.In this paper, we focus on how to remove shadow from a single color image and derive the representation of it in uniform light circumstance automatically. We project band-ratio vectors of image pixels onto illumination-invariant direction to obtain illumination-invariant grayscale image, then employ a linear color restoration mechanism to acquire shadow-free color image. Our method has three advantages. First, it runs very fast and realizes real-time process. Second, it has powerful adaptive ability and runs without any human-aid. And last, it reserves original circumstance information which is very important for image understanding.This paper has five chapters. In chapter two, we analyze several traditional image processing algorithms related to shadow removal, summarize their advantages and disadvantages, and propose our requirements and aims of shadow removal. In chapter three, we analyze the formation of color image and its color property, introduce theories of color constancy and color invariance, analyze many color invariant descriptors by experiments and choose band-ratio as characteristic quantity in our algorithm. In chapter four, we first interpret the cause of shadow formation based on RGB color model, then prove our idea by chromaticity, propose a method that projecting band-ratio vectors onto illumination-invariant direction to derive illumination-invariant grayscale image, at last get the shadow-free color image by a linear color restoration mechanism. We tested our algorithm by programming in Matlab and proved its effect and performance. Then we programmed with C++ language based on OpenCV library and integrated it into our reading robot system as a module in order to strengthen the application capacity of the system in a complex light circumstance. In the last chapter, we summarize the advantages and disadvantages of our algorithm, and point out how we can improve it.