Advanced Contour and Shape-Adaptive Still Image Compression Algorithms and Its Quality Assessment Method

As the existing image compression standards are mostly block-based and do not take the image contents into consideration, the compressed images appear to have blocking and ringing artifacts at high compression ratios. Therefore, in order to utilize image characteristics and achieve better visual qualities, we suggest to segment the image into several self-correlated regions in the first place and encode these regions separately. In this thesis, we first propose an efficient algorithm to encode the contours of regions from the image segmentation. The key feature of the proposed algorithm is using the DCT bases to approximate curve segments. In addition, we employ the weighted curvature measurement and the iterative split and merge calculation structure to divide the whole contour into suitable segments to be estimated piece-wisely. Comparisons with other existing contour approximation methods show the proposed method outperform other algorithms in either approximation error or compressed bits. In the second part of the thesis, we describe the proposed shape-adaptive image compression algorithm. Following the contour compression from part one, this method further employs the shape-adaptive versions of the image prediction, orthogonal transform, and image deblocking techniques. The resultant objective compression performance of our method is better than other block-based approaches and comparative to the JPEG2000. For the subjective comparisons, the proposed compression method has the best visual quality and more reasonable degradation at low bitrate. Moreover, to model the visual quality more accurately, we further propose an image quality assessment (IQA) method in the third part of this thesis. The proposed IQA method employs adaptive gradient and luminance similarity along with gradient-oriented pooling strategy. Comparing the shape-adaptively compressed images with the proposed IQA method, we find the results accordant to the subjective visual qualities. In addition, the performance on public image database also suggests the proposed method can be an alternate competitive IQA approach.