Image and Video Compression Techniques Based on Super-Resolution Reconstruction Algorithm
|School||Shanghai Jiaotong University|
|Course||Communication and Information System|
|Keywords||image/video compression super-resolution recon-struction motion estimation Kalman Filtering|
In many real-world application scenarios high-resolution images or videos areoften required while only low-quality images or videos are available due to the limitedbandwidth or storage. Therefore, the problem to reconstruct high-resolution versionsfrom the quality degraded sources has attracted many attentions. The super-resolutionreconstruction is recognized to be one of the most promising ways to overcome thisquality-limited problem.Conventional image coding schemes have poor performance in the case of lowbit-rate coding. Limited by the available resources, low bit-rate image coding has at-tracted tremendous interest in recent years. Motivated by recent research in computervision and super-resolution reconstruction technologies, this paper proposes an imagecompression scheme, in which we incorporate a learning-based image super-resolutionreconstruction algorithm into the mainstream image compression framework. Specif-ically, the super-resolution reconstruction process is regularized by the prior manifoldonly on the primitive patches. Each primitive patch is modeled by a sparse repre-sentation concerning an over-complete dictionary of trained set. Due to low intrinsicdimensionality of primitives, the number of samples in the dictionary can be great-ly reduced. Considering the similar geometry of the manifolds of the feature spacesfrom the low-frequency and the high-frequency primitives, we hypothesize that thelow-frequency and its corresponding high-frequency primitive patches share the samesparse representation structure. In this sense, primitive patches in the high-frequencylayer can be synthesized from both the high-frequency primitive patch dictionary andthe sparse structure of the corresponding low-frequency primitive patches. Only thedown-sampled image will be encoded and the super-resolution reconstruction algo- rithm is utilized to recover the high frequency information which has been removedby down-sampling. Experimental results show that our scheme achieves better objec-tive visual quality as well as subjective quality compared with JPEG2000 at the samelow bit-rates.For vision-based compression on low-quality data, this paper proposes a genericvideo compression framework based on video super-resolution technologies and mo-tion estimation using Kalman filtering. In our proposed video compression framework,a few selected frames (reference frames) are encoded at normal resolution while theother frames are encoded at reduced resolution. At the decoder, video super-resolutiontechnologies can be used to help the up-sampling and recovering processes of thosedegraded frames. However, the super-resolved high-resolution frames may not bepiecewise smooth and consistent with the neighbor reference frames. We are inter-ested in using motion information to enhance the spatio-temporal variation regularity.As we all know, highly accurate general motion estimation is not available in SR al-gorithm and inaccurately estimated motion often leads to disturbing artifacts. Kalmanfilters are utilized to obtain an optimal estimation of motion vector. In detail, we usethe motion information of the low-resolution level as the measured value and the pre-dict system equation to estimate the real motion vectors (state value). The relatedpaper“Super-Resolution Reconstruction with Prior Manifold on Primitive Patchesfor Video Compression”has been honored Top 10% Paper Award as the proceedingsof MMSP2011.