Image And Video Compression Based On Image Structure Characteristics

With the fast development of electronic devices, computers, networks and thegrowing demand of people, digital image and video data are generated and spread withrather high speed. The resolution, frame rate and dynamic range of those images andvideos tend to be higher. The image types become more plentiful. Besides the cameracaptured natural images, computer rendered screen images and depth images becomepervasive. Storage and transmission of these image and video data require high efficientcompression methods to reduce the size of data while maintaining good image quality.Conventional compression schemes mainly make use of prediction, transform,quantization and entropy coding to reduce the data redundancy and achievecompression. Prediction exploits the spatial and temporal local correlations; transformexploits the low-pass characteristics of the image and can compact the energy to lowfrequency bands. However, these methods have not fully exploited the structurecharacteristics of images for compression. In this paper, we pay attention to the structurecharacteristics of images and videos, including the characteristics presented in local andnon-local regions. For the text and graphics contents of the compound image, weconsider the high frequency characteristic, the sparse color histogram characteristic andthe repeated pattern characteristic. For natural images and videos, we notice thepresence of the image structure similarity. For depth images, they usually have steepchanges around the object boundaries. Based on these characteristics, we proposedsome methods to exploit these characteristics and achieved significant coding gain. Themajor contributions of this thesis can be outlined as follows:1. According to the local characteristics of the compound image, we proposed twospatial coding methods: Residual Scalar Quantization (RSQ) method, Base Colors andIndex Map (BCIM) method. Text/graphics block usually presents complex geometryshapes and has multi-directions. The prediction residues over text/graphics blockusually show strong anisotropic features. Transform would spread the energy to middleand high frequency bands which makes the compression difficult. In RSQ method, weskip the transform and directly perform quantization and entropy coding over theprediction residues. This can give more compact representation of the residues. Inaddition, text/graphics block usually has very few dominant colors. BCIM methoddescribes such a block compactly by several base colors and an index map. We decide the optimal base colors which minimize the distortion using dynamic programming.Rate-distortion optimization is utilized to decide the optimal base color number. Wedesigned the two methods as two intra coding modes, which were easy to be integratedinto the block-based coding framework. Due to the effective exploitation of the spatiallocal structure characteristics of the compound image, the two methods achieve evenmore than10dB gain in terms of PSNR for compound image compression.2. The text/graphics contents in a compound image usually present similar patternsover non-local regions. Based on this characteristic, we proposed the using ofone-dimensional dictionary-based Lempel-Ziv-Markov Chain Algorithm (LZMA) tocompress the compound image. Through matching the coding strings from a maintaineddictionary and using compact ways to represent them, LZMA can effectively make useof the characteristic of repeated patterns to reduce redundancy. To obtain highcompression efficiency even for noisy/quantized text and graphics contents, we haveextended LZMA to support lossy dictionary-based compression. We developed it as anew intra mode and implement it into block-based coding framework. Experimentalresults show that the proposed scheme achieves significant coding gains for compoundimage compression.3. Natural images/videos present structure similarity over non-local regions. Basedon this characteristic, we proposed the signal dependent transform (SDT) to improve thecoding performance. There exist strong correlations in non-local regions of theimages/videos, which means that the contents with similar structures usually appearover non-local regions. However, it is rather difficult to make use of these non-localcorrelations while simultaneously minimizing the overhead. To solve this problem, wepropose the signal dependent transform, which is derived from the decoded non-localblocks that are selected by matching neighboring pixels. Since the encoder and decodercan use the same methods to derive the proposed transform, we can successfullyeliminate the overhead. We have implemented the proposed transform into the KeyTechnology Area (KTA) software to exploit both spatial and temporal non-localcorrelations. The experimental results show that the coding gain over KTA can be ashigh as1.4dB in intra-frame coding, and up to1.0dB in inter-frame coding.4. Kinect-like depth and color images/videos are simultaneously captured and theyare generally aligned (captured at the same time from near views). For the depth image,it usually presents steep changes over the object boundaries. Considering thesecharacteristics, we designed an object-based coding system for the depth and colorimages/videos compression. The separate and independent coding of the different object planes for the depth images/videos can avoid the inefficiency coding of edges blocks atthe object boundaries and thus brings obvious coding gain. Moreover, the attractivefunctionality of “content-based” coding which permits the transmission of the interestedobject planes rather than an entire image provides a practical way to decrease the bitratewhile maintaining the reconstruction quality over the interested object.Our research work focus on the development of efficiency compression methods byexploiting of the image structure characteristics. Based on the structure characteristicsof compound images which are different from those of natural images, we designedspatial domain compression methods to significantly improve their compressionefficiency. There are some structure characteristics not fully exploited over naturalimages/videos. Non-local structure similarity is one of such characteristic. Experimentalresults indicate the effective exploitation of the image structure characteristics can bringgreat hope for improving the compression performance.