G Qiu, J Guan, "Interactive Image Matting using Optimization", Report-VIPLAB-01-2006 , Visual Information Processing Lab, School of ComputerScience and Information Technology, University of Nottingham, January 2006

Abstract - In this paper, we formulate interactive image matting as a constrained optimization problem. We first make some simple and reasonable assumptions about the alpha matte and assume that, geometrically, the closer two pixels are, the more likely they will have similar alpha values, conversely, the farther apart two pixels are, the more likely they will have different alpha values; photometrically, the more similar two pixels are, the more likely they will have similar alpha values, conversely, the more different two pixels are, the more likely they will have different alpha values. We then formulate these assumptions in a quadratic matting cost function and obtain the alpha matte by minimizing the matting cost function. User interaction in the form of a few scribbles indicating a few definite background and foreground pixels is used to provide constraints to make the problem well posed. For a given set of constraints the matting cost function has a unique global minimum and can be solved efficiently using standard methods. With the computed alpha matte we then estimate the background and foreground pixels. Results show that the new method works effectively and provides an alternative computational algorithm for building interactive image editing tools..

J. Guan and G. Qiu, "Image Contrast Gain Control by Linear Neighbourhood Embedding",  Technical Report, Report-VIPLAB-02-2005, School of Computer Science and Information Technology, University of Nottingham, November 2005

Abstract -In this paper, we present a method that adaptively computes a contrast gain control map for the image through the use of a novel technique termed linear neighborhood embedding (LNE) which first computes a locally linear relation for each pixel and its neighbors and then embeds these relations globally in the gain map image. We borrow the “think globally fit locally” concept and computational techniques from locally linear embedding (LLE) and compute the gain control image in closed forms by solving constrained optimization problems. We constrain the gain map locally following a gain contrast control mechanism similar to that found in the visual cortex to ensure that weak local contrasts are boosted and strong local contrasts are compressed, and propagate these local constraints globally following the original image pixels’ locally linear relations. We have applied our technique to compress high dynamic range images for reproduction in low dynamic range media and to enhance ordinary digital photographs. Results demonstrate that our technique is capable of preserving local details while avoiding artifacts such as halo

G Qiu, J Fang, "Classification in an Informative Sample Subspace", Report-VIPLAB-01-2005 , Visual Information Processing Lab, School of ComputerScience and Information Technology, University of Nottingham, May, 2005

Abstract -We have developed an Informative Sample Subspace (ISS) method that is suitable for projecting high dimensional data onto a low dimensional subspace for classification purposes. In this paper,
we present an ISS algorithm that uses a maximal mutual information criterion to search a labelled training dataset directly for the subspace’s projection base vectors. We evaluate the usefulness of the ISS method using synthetic data as well as real world problems. Experimental results demonstrate that the ISS algorithm is effective and can be used as a general method for representing high dimensional data in a low-dimensional subspace for classification.

G Qiu, J. Duan and G. D. Finlayson, "Adaptive dynamic range compression for high contrast images", Report-cvip-01-2004, March 2004

Abstract - In this paper we present a learning-based image processing technique. We have developed a novel method to map high dynamic range scenes to low dynamic range images for display. We formulate the problem as a quantization process and employ an adaptive conscience learning strategy to ensure that the mapped low dynamic range displays not only faithfully reproduce the visual contrast impressions of the original scenes but are also visually pleasing. This is achieved by the use of a competitive learning neural network that employs a frequency sensitive competitive learning mechanism to adaptively design the quantizer. The Optimization of an L2 distortion function ensures that the mapped low dynamic image preserves the relative contrasts of the original scene. The frequency sensitive competitive mechanism facilitates the full utilization of the limited displayable values. We present experimental results to demonstrate the effectiveness of the method in displaying a variety of high dynamic range scenes.

Abstract - In this paper, we present two novel, computationally efficient, practically easy to use and highly effective tone mapping techniques for the display of high dynamic range (HDR) images in low dynamic range (LDR) devices. The first new method, termed hierarchical nonlinear linear (HNL) tone-mapping operator maps the pixels in two hierarchical steps. The first step allocates appropriate numbers of LDR display levels to different HDR luminance intervals according to the pixel populations falling onto the intervals. The second step linearly maps the intervals to theirs allocated LDR display levels. In the developed HNL scheme, the assignment of LDR display levels to the HDR luminance intervals is controlled by a very simple and flexible formula with a single adjustable parameter. The second new method, termed the tree structured linear to equalization mapping (TS-LEM) HDR tone-mapping operator clusters the high dynamic range pixels in a hierarchical manner. A single parameter effectively controls the clustering in a computationally highly efficient way, which elegantly renders the mapping ranging from linear scaling to histogram equalization in a seamless manner. Compared with tone mapping operators involving spatial processing, our new methods are computationally more efficient and have far fewer parameters users have to set thus making them much easier to use. Compared with other tone reproduction curve based operators, our methods are much more flexible and effective. Experimental results further demonstrate that the performances of our techniques compare very favorably with state of the art while having the advantages of being computationally simple and practically easy to use. We also demonstrate that our new operators can be used for the effective enhancement of ordinary (8 bits per pixel/gray-scale or 24 bits per pixel/color) images.

G Qiu and J. Duan, "An Optimal Tone Reproduction Curve Operator for the Display of High Dynamic Range Images", Report-cvip-02-2004, February, 2004

Abstract - We present a new tone mapping method for the display of high dynamic range images in low dynamic range devices. We formulate high dynamic range image tone mapping as an optimisation problem. We introduce a two-term cost function, the first term favours linear scaling mapping, the second term favours histogram equalisation mapping, and jointly optimising the two terms optimally maps a high dynamic range image to a low dynamic range image. We control the mapping results by adjusting the relative weightings of the two terms in the objective function. We also present a fast and simple implementation for solving the optimisation problem. We will present results to demonstrate that our method works very effectively.