High Dynamic Range Image Exposure Fusion In Motion Scenes

Dynamic range means ratio between the maximum and minimum of apixel value in an image. In real world, the dynamic range is much largerthan two orders of magnitude, which is commonly seen in digital images.As a result, the high dynamic range image is introduced to expand thedynamic range of normal images, and to achieve better display quality.The generation of high dynamic range image can be concluded intotwo categories. The first one is to directly generate a raw high dynamicimage using images of different exposures. Then image which can bedisplayed on normal CRT is produced with the help of tone mappingalgorithm. This kind of method requires estimation of camera responsecurve which evidently introduces nonlinearity. This may causeunsatisfactory results. The other method is to skip the step of generatingraw high dynamic image. Instead, an image which can directly be shownon CRT is generated using exposure fusion. This article is based on thesecond class of method. As a result, non-linear operation is not introduced and algorithm complexity is lower than the other method.Although the method mentioned above works well for static scenes,things are different when it comes to the situation in which camera shakeand moving objects exist. In order to eliminate possible ghost effect, thisarticle will introduce two algorithms aiming at saturated and non-saturatedregion of a reference image. In addition, an image copy generationalgorithm is also used to compensate for positions where pixels are judgedas ghost pixels, in order to enrich the image details.For non-saturated region, the image will be split into blocks. In eachblock, reciprocity law is used to judge how many pixls in this block areghost pixels. If number of ghost pixles exceeds certain threshold, thisblock will be judged as ghost block and abandoned. As soon as we getstatic scene from the image, a brightness transfer function can beestimated between reference image and images with different exposures.With the help of this function, the ghost pixels in saturated region can alsobe detected. A threshold is also set to decide if one pixel in referenceimage is ghost pixel after transferring to target image using brightnesstransfer function.In order to gather image information from image set as much aspossible, the region which is detected to be ghost regions should not be easily set to zero in weighting function. An algorithm which can produceimages of different exposures with same content from reference image isintroduced. It will replace the region with ghost pixles and improve thequality of fusion results.Comparing with other algorithms dealing with motion scenes, ouralgorithm has already achieved satisfying results. But an optimizationalgorithm is also proposed in this article. With the help of gradientenhancement, the image sharpness can be further improved. Using the factthat big gradient represents image structure and small gradient representsimage details, we decide to attenuate big gradient and enhance smallgradient. As a result, the details of an image are enriched and visual effectof the image is increased. Afterwards, the original image is reconstructedusing a possion solver. Experiments results show that the reconstructedimage outperforms original image both in image sharpness and localcontrast.