Learning-Based Restoratioin For Compressed Images

With the rapid development of the digital multimedia technology, there is a variety of growing demand for the digital image. Since the digital image has the characteristics of a great amount of information, it is required to be compressed before transmission and storage. At present, Block-based discrete cosine transform (BDCT) coding has prevailed in the mainstream image and video compression standards, which aims at reducing the inter-pixel statistical redundancy. However, for the sake of achieving higher compression ratio, BDCT together with the coarse quantization has caused too much lost of the original information, and this gives rise to the discontinuity of intensities between adjacent blocks which are named as blocking artifacts, and truncates the high frequency (HF) DCT coefficients, which results in ringing artifacts around the contours. Therefore, the visual quality is unsatisfied on both the subjective and the objective quality, so effective postprocessing scheme for compressed image is urgently demanded. In this paper, in order to improve the subjective image quality, we carefully researched the de-blocking algorithm and proposed two learning based image restoration algorithm, which efficiently remove the blocking effect and improve the subjective quality of the image. (1) Compressed image restoration based on Field of Experts: this scheme combines the popular sparse coding method with Markov Random Field. On one hand, we learn the prior knowledge from a set of natural images, and modeled it as a high order Markov Random Field based on the Field of Experts framework. On the other hand, according to the features of the quantization noise of compressed image, a degradation model, represent by additive Gaussian noise model, is proposed to simulate this process. Finally, we can get the restored image after several iterations by using the maximum a posteriori criterion. (2) K-SVD based compressed image restoration: this method depends on a novel adaptive dictionary learning method——K-SVD method, by using sparse coding, sample image patches are represented with the dictionary learned by this K-SVD. During the image restoration process, we should extend this sparse coding method from small image patch onto the whole image. Here we create a proper objection function and solve this function by using an appropriate updating method, after several iterations, a reconstructed image with high quality, which satisfies the human visual system, could be obtained.