A ROBUST AND FAST ANTI-GHOSTING ALGORITHM FOR HIGH DYNAMIC RANGE IMAGING Shiqian Wu, Shoulie Xie, Susanto Rahardja and Zhengguo Li Institute for Infocomm Research, 1 Fusionopolis Way, #21-01 Connexis Singapore 138632 Email:{shiqian,slxie,rsusanto,ezgli}@i2r.a-star.edu.sg ABSTRACT This paper presents a robust and fast algorithm for automatically generating high dynamic range (HDR) images in presence of camera movement and moving objects. This scheme comprises five modules: 1) image alignment, 2) estimation of camera response function (CRF) in dynamic scenes, 3) moving object detection, 4) progressive image correction, and 5) construction of HDR images. The key advantage of the algorithm is the ability to generate HDR images without ghost artifact. The proposed algorithm is fast as it is a one-shot solution without iterative computation and post-processing or even manual operation. Experimental results demonstrate that the proposed method outperforms the existing commercial products. Index Terms—HDR imaging, ghost artifact, camera response function, detection of moving objects, progressive image correction 1. INTRODUCTION Many natural scenes have higher dynamic range than those recorded by a camera. Combining differently exposed low dynamic range (LDR) images of the same scene has been the most popular approach to generate an HDR image [1, 8]. However, these methods always result in two problems [2, 3, 4, 8]: 1) Ghost artifact appears due to dynamic scenes, such as moving people, vehicles and so on; 2) Blurry artifact occurs because of camera movements. While image registration has been employed to deal with the camera movement, ghost removal remains an open issue in HDR image generation. Recently, statistical methods were employed to generate HDR images without ghost artifacts in [4] and [7]. These techniques do not rely on explicit object detection and motion estimation. But expensive computation is unacceptable for a large set of images due to its iterative strategy. It is also noted that the statistical methods cannot completely remove ghost artifacts unless the probability of a pixel belonging to the background is absolutely zero or one. Moreover, as indicated by these authors, their methods are exposed to getting halo effects in the generated HDR image, especially in image sequences including wavy motion like burning candle flames. In [2], Grosch presented an alternative approach for movement detection based on the error map, where the error map was resulted from the predicted image via CRF in a specific exposure and the original image under the same exposure. This approach works well if moving objects are not in saturated or underexposed regions. Due to inaccuracy of the CRF, the error map is non-uniform, i.e., the predicted error is big for bright regions while small for dark region within two LDR images, and the predicted error increases along with the increase of exposure gaps. Accordingly, this method cannot thoroughly identify the moving objects so that the regions of the moving objects have to be corrected with a special clone brush. Moreover, the threshold is not robust as it depends on the amount of noise in the images. In [8], movement is detected based on the variation of variance. But this technique is sensitive to the estimated error of CRF, and this method is only effective to remove ghost artifacts which can be segmented. The concept of entropy was proposed to identify moving objects in [3]. Although entropy is independent of image intensities, this method can only cope with the scenes in which ghosts occur in regions with low dynamic range. In this paper, a fast and robust algorithm is developed for automatic generation of HDR images in consideration of camera moment and dynamic scenes. Five criteria are first proposed to estimate the CRF in dynamic scenes. Three criteria are proposed to identify moving objects. An idea of progressively updating strategy is employed to identify the difference between moving objects and background in LDR images. A novel algorithm combining the CRF information and the original images is presented to correct the LDR images. To mitigate the errors from CRF and image alignment, the corrected images are further processed using image inpainting. The proposed method is non-iterative. Also, the approach is robust to different images (with different contents, dynamic scenes/still scenes) without tuning the predefined parameters, i.e., the thresholds are robust to varied scenes. The experimental results demonstrate that the proposed method is significantly superior to the existing commercial products [9~11]. 397 978-1-4244-7993-1/10/$26.00 ©2010 IEEE ICIP 2010 Proceedings of 2010 IEEE 17th International Conference on Image Processing September 26-29, 2010, Hong Kong