Implementation of Retinex Algorithm by Eyegaze Tracking Interface Ryo Ohtera, Takahiko Horiuchi, Shoji Tominaga Graduate School of Science and Technology, Chiba University, 1-33 Yayoi-cho, Inage-ku, Chiba, 263-8522, Japan ABSTRACT This paper proposes a computer vision system for improving the image quality around a steady gaze point on a display device. We assume that one observes a localized small area of the displayed image, rather than the whole image, because of a limited visual angle. The computer vision system consists of two subsystems which are an eyegaze detection subsystem and an image quality improvement subsystem. The eyegaze detection subsystem tracks a human gaze point on the display. A tracking algorithm is developed for capturing a human face from a single monocular camera without using any special devices. The image quality improvement subsystem performs a localized Retinex algorithm. Although the conventional algorithms contain a large number of complex computations, the localized algorithm is devised for performing the Retinex computation in high speed for only a localized part within the whole image. The combined system is developed so that the image quality is improved in real time within just local region around the detected gaze point. We make an experimental system consisting of an off-the-shelf digital video camera and a personal computer. The whole performance of the computer vision system is examined experimentally on subjective assessment and processing time. Keywords: Retinex model, eyegaze tracking, human interface, image quality improvement 1. INTRODUCTION We are much concerned about the computational cost in applying some image processing algorithms for the purpose of improving image quality. When we watch an image on a display monitor, our visual angles are usually limited, and we observe a localized small area of the displayed image, rather than the whole image. In such a case, the image processing is not necessarily applied to the whole image, but can be applied to only the local region around a steady gazing point by the observer. An approach to localizing the region of image processing is very effective to reduce the computational cost. The approach has an advantage that the computational cost is independent of the size of an input image. The Retinex image enhancement algorithms, originally proposed by Land [1]-[5] and McCann [6], provide an automatic image enhancement method for improving the image quality of digital images. As is well-known, the human visual system has the visual ability of lightness constancy and color constancy. So, one can understand a natural scene in the world without the effect of illuminant with spatially non-uniform distribution. The Retinex computation method has a long history for estimating the illumination component from the image captured by a camera. The Multi-Scale Retinex (MSR) [6]-[11], which integrates multiple Single-Scale Retinex (SSR), can suppress the unwanted halo artifacts around high contrast edges. However it takes high computation cost, compared with the SSR. In addition, the computation cost depends on the size of an input image. Once we considered a method for reducing the computation cost by developing a serial Retinex algorithm for time-sequential processing [12]. However, further reducing the computation cost is required for applying the algorithms to real situations. In this paper, we consider a unique approach based on the eyegaze interface for improving the computational efficiency. A computer vision system is proposed for improving the image quality by applying the Retinex algorithm to the local region around a steady gaze point on a display device. It is expected to lead to solving the problem of high computation cost. The proposed computer vision system can realize the process, and the system consists of two subsystems which are an eyegaze detection subsystem and the image quality improvement subsystem. In the eyegaze detection subsystem, we use a simple eyegaze detection method [13] using the parametric template matching[14] to detect a human gaze point. In the image quality improvement subsystem, we develop a localized Retinex algorithm. Although the conventional algorithms contain a large number of complex computations, the localized algorithm is devised for performing the Retinex Color Imaging XIV: Displaying, Processing, Hardcopy, and Applications, edited by Reiner Eschbach Gabriel G. Marcu, Shoji Tominaga, Alessandro Rizzi, Proc. of SPIE-IS&T Electronic Imaging SPIE Vol. 7241, 724107 · © 2009 SPIE-IS&T · CCC code: 0277-786X/09/$18 · doi: 10.1117/12.805797 SPIE-IS&T/ Vol. 7241 724107-1