Research On Image And Video Compression Algorithm Based On Non-local Sparsity

Wireless multimedia sensor networks (WMSNs) collects multimedia data by lots of image, video or voice sensors, and then transmits them to the gateway node and sink node. The emerge of WMSNs has given us a wide range of applications, such as in-telligent transportation,intelligent surveillance and smart city. Therefore, it is widely studied by academy and industry. Due to the rich data types and huge amount of data, it needs to be compressed before transmitted. Although multimedia data compression standard has an unique performance, it gives a large burden to the encoder of a node. Besides, WMSNs’nodes have the restrained resources, including energy restrain, mem-ory restrain and calculating ability restrain. Hence, huge multimedia data compression has already been a challenge for WMSNs. Compressed sensing is a new sampling the-ory, which just needs a small amount of data to reconstruct the original data. On the one hand, compressed sensing can avoid the restrain of the Nyquist law and reduce the amount of transmitted data, thus it reduces the burden of memory space and network transmitted ability. On the other hand, at the encoder compressed sensing just needs to do nonlinear projection. Thus it reduces the demand of calculating ability of nodes. Therefore, compressed sensing is very suited for the WMSNs. In the structure of com-pressed sensing, the most important aspect is reconstruction algorithm. Along with the study of reconstruction algorithm, researchers come to realize that model-based recon-struction algorithm can bring us the increased performance.In this paper, we consider to exploit the new model, i.e., non-local sparse model (NLM), to the compressed sensing of image and video for the WMSNs. NLM exlpoits the correlation between similar patches, which widely exist in the image. It is a realise model of image that can help us obtain good performance. When applying NLM to the compressed image sensing, we not only exlpoit the NLM to obtain sparser representa-tion, but also deem the images as the compressed signals to get more details of images to recovery. Firstly, we propose to group similar patches into3-D groups. By using3-D transform, we can efficiently exploit the correlation between similar patches. When getting the sparse samples, the weighted matrix of empirical wiener filtering is used to not only retain big coefficients, which represents sparse samples, but also retain small coefficients, which represents details.When applying NLM to the compressed video sensing, due to video is the excep-tion of multiview video, we select multiview video to study. In the research, we have two inovative points. One is to exploit NLM to remove the three redundancy and an-other is to use more information provided by multiview video to retain more details of images. Firstly, NLM can help us remove the three redundancy easily. It just needs to find the sufficient patches including patches from different moment but same views and patches from different views but same moment. To make sure that the similarity between patches is accurate, we propose a joint match rule of similar patches. Based on previous work, joint filtering on the group of similar patches is proposed. Since self-adaptive base can bring us spareser representation than fixed base, we design a convex problem for the joint base of similar patches. Meanwhile, we exploit empirical wiener filtering to obtain weighted matrix of similar patches. Therefore, we can gain not only higher PSNR but also better details of images.At last, lots of experiments are utilized to verify the performance of our proposed algorithms. It notes that our algorithms have more advantages than mainstream algo-rithms both in the PSNR and in the visual quality.