Liu, Bo-zhi
(2018)
Understanding colour image: colour constancy.
PhD thesis, University of Nottingham.
Abstract
Human visual system has a mechanism which ensures that the perceived colour of an object remains almost constant under varying illumination conditions, and this mechanism is called colour constancy. Electronic imaging systems such as digital cameras do not naturally have this ability. The color appearance of images of an object under different lighting conditions changes with the colour of the light sources and this can cause problems in many computer vision applications such as object recognition. To deal with this problem, many algorithms have been developed to estimate the input image’s illuminant, and then recover the intrinsic colour of the scene correctly. In this thesis, we focus on this topic, try to produce new colour constancy algorithms in both images and videos, to improve the performance of the state of the art.
This thesis makes four technical contributions. First, we have developed a new image representation scheme suitable for developing learning based colour constancy algorithms; second, we introduce a new method that formulates the colour constancy problem as one that infers the illuminant class of the input image; third, we introduce a novel clustering classification colour constancy framework (the 4C method); and finally, we extend our method from still image into video processing, create a new framework to deal with the colour constancy problem in videos.
As in many computer vision problems, one of the crucial issues is how to effectively represent the input events. Colour constancy is no exception and we need to first represent the input image. As we are only interested in the colours of the image, colour histogram is a natural choice. However, traditional colour histogram is content dependent. As our task is estimating the colours of the illuminant rather than the colours of the image, we need a representation that is relatively independent of the image content. Based on this reasoning, we introduce the novel concept of a binary colour histogram where it records if a colour has appeared in the image or not and disregards the frequency of the colours appear in the image. We will present experimental results to demonstrate that our new binary histogram representation is particularly suitable for learning based colour constancy and that it provides better performances than other traditional representation schemes.
The colour of a digital image is directly affected by the colour of the illuminant. We reason if we can recognize or classify the illuminant source of the image, we can then correct the colour of the image. Based on this rationale, we formulate the colour constancy problem as an illuminant classification problem. We assume that each image has an associated class of illuminant and the task of colour constancy is that of recognizing the illuminant class of the image. To accomplish this, we make use of our newly introduced binary colour histogram representation scheme and employ a powerful machine learning method called the Random Forest to construct the illuminant recognition system. We will present experimental results to show the effectiveness of our new method.
Encouraged by the success of our illuminant recognition framework, we have developed a novel clustering classification colour constancy (the 4C) framework. We reason that similar illuminants will result in similar white point colours in an image. Based on this assumption, we first use a clustering algorithm to group similar white point colours of the training samples into the same cluster. We then treat the images in the same cluster as belonging to the same illumination source and each cluster as one class of illuminants. The colour constancy problem, i.e., that of estimating the unknown illuminant of an image, becomes that of identifying which illuminant class (cluster) the image’s illuminant falling into. We again make use of our novel binary colour histogram representation and our random forest based illuminant classification methods to implement our new 4C colour constancy framework. We present experimental results on publicly available testing datasets and show that our new method is competitive to state of the art.
As a practical application, we have successfully extended our novel colour constancy methods from still image into video processing. The video tonal stabilization problem is still an unsolved problem, and current algorithms are only focusing on keeping the tonal stable during video playing, not really trying to recover the incorrect illuminant. We tackle these two problems together by keeping the tonal stable and recovering the frame colour to a canonical illuminant. Our approach first divides video frames into shots containing similar illuminant characteristics. We then correct the frames in the same scene by using the Random Forest illuminant estimation framework. A smooth function is applied to prevent flick and flash from occurring at the boundary of the neighboring scenes. Experimental results show that our new methods can improve video quality effectively.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)