Content Based High Performance Intra Coding Study

Independent decodability enables Intra coded frame to be used as random access point and preventer of error propagation, which further makes intra frame very essential to various video applications. However, the relatively low compression ratio of Intra coded frame and its importance formed an urgent conflict which needs to be solved. The recently proposed Intra coding technologies are mostly based on the hybrid intra coding framework adopted by H.264/AVC. Therefore, understand and perfect how the coding theory is applied to the hybrid intra coding framework and propose advanced intra coding algorithms based on the study of the theory is of substantial theoretical and practical significanceFor theoretical analysis, this work perceives that the hybrid intra coding framework break down the entropy of a frame into the sum of the conditional entropy of many successive pixel blocks according to the chain rule. To reach the conditional entropy of the blocks, Intra Prediction is introduced to remove the mutual information of different blocks. When coding each block itself, Transform is applied to map the vector composed of pixels within the block to a vector composed of independent symbols in a transform domain. Finally, entropy coding is applied to each symbol in the transformed domain to reach its entropy independently. Since it's hard to really express and deal with mutual information, we usually take linear correlation as an approximation of mutual information to be used as the optimization goal. This work analyzed the relationship between mutual information and linear correlation as well as the correctness of the approximation.As an extension of the theoretical analysis at the abstract level, this work studied how to generate the optimal single mode joint Intra Prediction and Transform for a specific region of image. In fact, the optimal Intra Prediction is the Wiener Filter that totally removes the linear correlation among reference pixels and the coded pixels, moreover, the jointly optimal Transform is the KLT for the prediction residues got from the optimal prediction.Based on the theoretical analysis, this work further focused on Intra Prediction and proposed 3 types of advanced prediction schemes which all applied the algorithm of adaptively generating the optimal intra prediction mode for a specific image region. The 3 types of schemes are suitable for different application scenarios. The normal adaptive intra prediction scheme has to transmit extra information of the generated prediction filters therefore is mainly used for high resolution sequences with relatively complex texture. The decoder derived adaptive intra prediction scheme avoids transmitting the overhead thus can be used for a wider range of resolution, yet it introduces much calculation complexity on the decoder side. The fixed prediction filter scheme generates prediction filters by off-line statistics thus has lowest complexity both in encoder and decoder side, yet its performance is largely affected by the statistical similarity between training sequence set and the coded sequence. Experiments showed that under the suitable application scenarios, all 3 schemes achieved significant coding gain compared with H.264/AVC, especially for high resolution sequences with complex texture.