The Design And Implementation Of Face Super-resolution Software In Surveillance Videos

This topic is derived from the ministry of science and technology benefiting the people of science and technology plan projects(2013GS230301),Qiqihaer city 3G-WIFI city safety monitoring and control system for promotion. With surveillance cameras and security monitoring systems being popular in social security and personal security field, more and more incidents which endangered public safety have begun to take surveillance video as an important evidence and an important source to obtain relevant information, but there are kinds of factors that affect the quality of the security monitoring systems in real life, which reduces the accuracy of the information. To solve this problem, this thesis take a deep research on the face super-resolution reconstruction algorithms in surveillance videos. From the perspective of the image super-resolution reconstruction, we study and improve the existing algorithms, and also introduce OpenMP parallel processing framework to build a face super-resolution algorithms display platform. The specific content are as follows:Firstly, we proposed the PCA-based face selection algorithm. Since the learning based super-resolution method always require huge sample libraries as support, which leading to the traditional algorithms run slowly and inefficient. Aiming at this shortcoming, we use principle component analysis to select the samples globally based on similarity, and remove the dissimilar faces in the library, only leaves a small amount of samples closest to reconstruction. The corresponding experiments showed that as long as the number of selected samples can be properly chose, this method can significantly reduce the library while ensure the same reconstruction quality. This method can also reduce the running time of entire algorithm.Secondly, we take a deeply study on sparse representation theory and the SR based face super-resolution method. We introduced and implemented the LCR and LSR face super-resolution algorithms. By analyzing the experiment results we found their results are extremely unstable when the input faces are under noise polluted environments or the real surveillance scenarios. Also the strict dependency on the patches positions reduces its robustness against position.Thirdly, aiming at the traditional SR based reconstruction methods' drawbacks, an improved algorithm based on global similarity selection and local similarity representation is proposed. Instead of sparsity, we use similarity to represent the relationship between patches, as well as taking the neighbor patches into account the representation, a novel two layer representation framework is built which has a natural parallel structure to implement easily. Through experiments under different noise environments, it showed this improved method has a great improve on the robustness against noise, compared to the traditional algorithms.Finally, under Visual Studio 2012 environment and OpenMP parallel processing framework, the traditional method of LSR, LCR and improved method are implemented. We also build a face super-resolution software platform according to the advise from Qiqihar Public Secure Bureau. This platform is capable of reconstructing faces under experimental environment and the real surveillance scenarios, and the corresponding parameters during the reconstruction process, such as PSNR, SSIM and time consuming can be also outputted, which achieved very good results.