| The brightness range of real scene and which human eye can distinguish is up to over 105, but the dynamic range of normal image sensor is no more than 103. In order to reproduce high dynamic range scenes, the same scene is multiple exoposed to obtain images of different exposures, and then these images are merged to get a high dynamic range image. However, dynamic range of display is mostly about 102, high dynamic range images can’t be displayed on normal display. It is needed to compress the luminance values. How to retain rich details and natural visual effects of the original scene has been a hot research in this field. In recent years, as the high dynamic range imaging technology is increasing used in consumer electronics, remote sensing, security monitoring, digital television and other fields, it has a strong practical meaning to research this topic.High dynamic range (HDR) image displaying technology mainly consists of two method. The first method is fusing multiple images of different exposures directly. And the second one is compressing the luminance values with tone mapping algorithm after obtaining a high dynamic range image by merging different exposed images. In this dissertation, two methods are studied, and the detailed work is as follows:Firstly, analyze the shortcomings of current multiple exopusure image fusion algorithms, and point out that correction method of initial fusion weights is the key to affect fusion performance. The content aware factor is introduced to the original guided filter in order to correct the initial fusion weights. For regions which have high visual saliency and rich edge information, the fusion weights have high fidelity. Experimental results show that sharp edges are preserved and image resolution is increased in the blending area of image, which is processed with the proposed algorithm.Secondly, propose a merging algorithm which corrects the overflow of transition region, and adopt winner-take-all strategy to reduce quantization error of merging images. The intensity of differently exposed images can be mapped by using linear fitting without knowing the exposure time. According to the overflow of neighbor pixels, establish a mechanism of classification correction, and the transition region of image is smooth.Thirdly, combined with the advantages of global and local operator, propose a tone mapping algorithm with automatic parameter adjustment. Use step-wise approximation method in order to achieve the best display brightness and extract the details of the image using a median filter. The proposed algorithm can effectively compress the dynamic range of HDR images and preserve details and local contrast as well. Compared with other algorithms, the poposed algorithm has a great advantage of image naturalness.Fourthly, considering the practical need of image signal processor (ISP), propose a simplified hardware design of merging and tone mapping algorithms. Storage local fitler data using line buffer; use pipelined design for complex combinational logic operations; improve pipelined design of 5×5 median filter, therefore, reduce the number of comparators from 360 to 125; use table-driven algorithm for approximating the result of exponential function, reducing the complexity greatly.Fifthly, simulate the design in RTL level. The results of Matlab floating point and Modelsim fixed point is consistent. Build FPGA verification platform by using Xilinx xc6slx45 device. The results of FPGA logical synthesis shows that this design can meet real-time video processing requirements of 720p resolution and 60fps frame rate. The software testing reults show that the performce of the design is good. |
PDF Full Text Request