Subband Architecture And Gradient Based Multi-exposure Fusion

Multi-exposure fusion method focuses on generating a single high quality resultant image by blending a set of multiple exposure images of a scene. The objective of multi-exposure fusion is representing all detail and color imformation of the scene naturally and authentically. Compare with HDR imaging, this method skips recovering the camera response curve stage and can be excuted conveniently.Multi-exposure fusion method can be used in varity fileds, such as image editing, high quality imaging in portable equipments like cellphone, enhancement of medical images, etc. Additionally, the applications of this technology in video area are on the horizon. For instance, the high dynamic range video camera for road memory driving record is on the market. This technology compensates the limited dynamic range of consumer-level digital camera and display devices which is narrower than real world scene's. It also makes obtaining professional high quality images only via consumer-level camera available.The thesis advances a novel multi-exposure fusion algorithm, which is based on subband architecture and gradient. Experiments demonstrate that this algorithm recovers detail imformation efficiently and accurately both in static and dynamic scenes. It devotes to unartificial result, meanwhile, strengthens the visual impact of the image. The main contributions of this thesis are summarized as follows:(1) An effective multi-exposure fusion framework is proposed. We use symmetrical analysis and synthesis architecture in which the analysis filters are the same type with the synthesis filters. The gain control maps are used for controlling the strength of the subband images which are obtained by decomposing the image with the analysis filters. We use weight maps to assign each pixel's contribution to the fusion result. Finally, the result is reconstructed by using the fused subband images with synthesis filters.(2) A novel method of controlling the subband images is presentd. The gain control maps are calculated according to the activity level of subband signals and can be used for controlling the strength of subband signals to avoiding nonlinear distortions caused by signal decomposition. The low and high frequency artifacts can be introduced by nonlinear processing. The synthesis filters can remove the low frequency artifacts but high frequency artifacts since they are tuned to the same frequency band as the analysis filters. The gain control strategy makes a compensation of it.(3) A robust motion detection method for exposure sequence is put forward. This method is based on the truth that the direction of gradient is insensitive to the change of exposure. We divide image into three levels, blank level, response level and object level. Every pixel only belongs to one level of them and is classified according to the votes got from its friend images. The friend images of one input are the rest images of the sequence except itself.