The Research On Image Compression Algorithm Based On Object Segmentation

Because of the limited memory storage of IC card, the certification photo requires a low bit rate while highly efficient coding algorithm. However, adopting traditional compression algorithms only lead to an unsatisfied compression results due to lack of awaring the unique characteristic of certification photo, that the potential target object is a human head figure with pure background color.In order to solve this problem, this thesis put forward the image compression algorithm based on object segmentation, which eliminates the background redundancy by coding human head figure and background respectively, through in-depth research on the characteristic of the certification photo. This algorithm mainly consists of three parts. One is the contour lossless coding of the arbitrary shaped object, the second is the texture lossy coding of arbitrary shaped object, and the third is the coding of the background color. Through this kind of division, it not only guarantees the fidelity of the edge of human head figure, which enhances the human subjective visual effect, but also makes it no need to do finer coding towards the pure-colored background. Consequently more bits can be allocated to describe the texture information of human head figure. Finally it could produce reconstructed certification photo with high human subjective quality and low bit rate.This thesis is started with the introduction of common image compression techniques and algorithms. After an analysis of the characteristic of certification photos is given, the compression algorithm of certification photo based on object segmentation is derived. Then a careful research is given toward lossless contour coding based on difference chain coding, shape-adaptive integer wavelet transform and shape-adaptive SPIHT algorithm. What's more, the complete flow chart of the whole algorithm is presented and described as well as the file format of the final bit stream. Finally, a number of experiment results show that under the conditions of bpp=0.078, our algorithm outperforms 4dB in average PSNR when compared with standard SPIHT algorithm.