Image Processing Towards Cloud Media

In recent years, multimedia applications and social networks are developing rapidly. Multimedia data on the internet is experiencing explosive growth. This brings new opportunities and challenges to image processing and computer vision community. How to utilize the huge database in the cloud and leverage them to advance related applications in image processing is becoming a hot research topic in academia.To explore the potential assistance of the huge database in the cloud to low-level image processing problems, we begin with exploring the correlations among cloud images. We propose a correlation modeling framework by combining global alignment and local patch matching. A matching scheme for degraded images(e.g. noisy images, low-resolution images) is proposed. We investigate how to process the classical image processing problems, such as image compression, super-resolution and denoising, using the proposed matching scheme. The proposed schemes achieve significant improvement compared with state-of-the-art methods. Our work provides a novel idea and new path for vision computing in the cloud era. The main contributions of this thesis are summarized as follows.1. Proposing image compression towards cloud media. It no longer compresses images pixel by pixel and instead tries to describe images by local descriptors and thumbnail images. In the encoder side, to efficiently compress high-dimensional feature vectors of local descriptors, we propose predictive coding which leverages its correlations with the thumbnail image. In the decoder side, the high quality image is instantly reconstructed from a large scale database by retrieving correlated images with local feature and stitching corresponding regions with the guidance of thumbnail. Experiments show that the proposed scheme could produce visually pleasing results even at thousands to one compression ratios.2. Proposing image super-resolution towards cloud media. Our scheme generates an adaptive, highly-correlated high-resolution image dataset based on the local descriptors of the observed low-resolution image. We propose a novel matching method to combine the high-level matching(image registration) and low-level patch matching. In low-level patch matching, we propose changing patch size according to the alignment accuracy. Experiments show that the proposed scheme achieves significant improvement compared with four state-of-the-art methods.3. Proposing image denoising towards cloud media. This thesis proposes a graphcut based patch matching scheme to overcome the matching difficulties in noisy image.We propose a two-stage based hybrid filtering method, which explores both internal and external correlations. The denoising result at the first stage is used to improve image registration, patch matching and estimation of filtering parameters in the second stage, which further improves the denoising result. Experiments show that our scheme achieves more than 2 d B gain compared with BM3 D at a wide range of noise levels.