Aiming at problem of additive Gauss noise removal, an improved image restoration algorithm based on the existing K-means Singular Value Decomposition (K-SVD) method was proposed by integrating dictionary learning and variational model. Firstly, according to geometric and photometric information, image blocks were clustered into different groups, and these groups were classified into different types according to the texture and edge categories, then an adaptive dictionary was trained according to the types of these groups and the size of the atoms determined by the noise level. Secondly, a variational model was constructed by fusing the sparse representation priori obtained from the dictionary with the non-local similarity priori of the image itself. Finally, the final denoised image was obtained by solving the variational model. The experimental results show that compared with similar denoising algorithms, when the noise ratio is high, the proposed method has better visual effect, solving the problems of poor accuracy, serious texture loss and visual artifacts; the structural similarity index is also significantly improved, and the Peak Signal-to-Noise Ratio (PSNR) is increased by an average of more than 10%.

Aiming at the complex communication environment and low efficiency of big data processing in agricultural Internet Of Things (IOT), a big data processing mechanism was proposed based on the adaptive collaborative opportunities. According to the requirements of agricultural application and the impact of agricultural environment on wireless data transmission, a cross-layer interaction analysis model was established, which was combined with opportunities for collaborative mechanisms and big data processing requirements. Then the design of a large data processing was proposed. Experimental analysis and testing show that the proposed big data processing scheme has better system throughput, reliability, and system processing performance than traditional coordination mechanism and data processing programs.

As the conventional physical cluster system fails to cope flexibly with large-scale Internet applications, a comprehensive load balancing mechanism for cloud-based virtual cluster system was proposed. It first periodically collected CPU and memory usage, number of connections, and response time of all virtual machines and physical hosts, then calculated the weighted load of the physical hosts, and finally scheduled and assigned the task requests based on the calculated comprehensive load, thus could adapt to the complex, dynamic and variable computing environment. The experimental results show that, compared with other scheduling mechanisms such as Weighted Round Robin (WRR) and Weighted Least Connections (WLC), the proposed mechanism is delay optimal under heavy workload, and moreover, it can increase or decrease the number of Virtual Machines (VMs) dynamically to balance the server load usually within 5 seconds.

Based on cloud computings architecture and the actual demands of Tsinghua University, followed by utilizing the advanced OpenStack platform, adopting hierarchical design method, the TsinghuaCloud platform that could be used to perform integrated management on cloud resources was designed and implemented. The advantages and main required module functions of this system were analyzed. Focusing on the resource scheduling, a strategy based on task scheduling and load balancing was proposed. The experiment and analysis of the scheduling plan verify that the scheduling strategy can balance servers resource load on the basis of ensuring its service performance and execution efficiency, so as to make the cloud platform relatively stable.

The Classification and Regression Tree (CART) algorithm was introduced, and the application of the CART algorithm in spam filtering was discussed. Firstly, text messages of email samples were pre-processed, regular email and spam training sets were trained to establish the single classifier with the CART algorithm. Then a boosting based model which was combined with multiple CART classifiers was established. The comparison test results show that the multiple CART classifiers based on boosting has achieved better results.

With the development of the short message services, the characteristics and contents of the spam short Message are also changing constantly, the main problems that exist in the traditional short message filtering systems are that the characteristics and contents fail to be updated in time, which reduced the filter capability. This paper mainly utilized Na ve Bays advantage of rapid statistics classification and Support Vector Machine （SVM） incremental training characteristic in Spam Short Message filtering, and provided feedbacks to the online filtering sub-system in time in order to enhance the system-s self-adaptability. The experimental results show that this new method effectively deals with the above problems in the traditional spam short Message filtering systems．

The problem of mining hidden communities on heterogeneous social networks was investigated. On the basis of the bibliographical data, this issue solved community mining problem in heterogeneous relation, and extracted the relation chain. Therefore, it can help researchers do some research work.