The Research Of Face Image Reconstruction

In recent years,with the rapid development of information technology,video surveillance has become indispensable for public security.Facial information has universality and collectability features,so it gets more and more attention.In the square,park,stations and other places,the camera is often installed in a broad sense,far away from the face,so the detected faces are in low resolution,and size of faces could be very small,so that faces can't be identified.In addition,the image processing is affected by many factors,such as illumination,motion,noise,and so on,for these reasons,the faces image quality is reduced,and the resolution is poor.The reconstruction method of small face image in video is researched in this thesis,including the key technologies such as face detection and tracking,single frame and multi frame face reconstruction method,to improve the face image resolution and identification in video surveillance.In addition,this thesis designed and implemented a kind of software for face detection and reconstruction.The main works of this thesis are as follows:In terms of face detection and tracking of video,face tracking method fused AdaBoost algorithm and CamShift algorithm is implemented.First,the face region detected by AdaBoost algorithm is used as the initial tracking search window of CamShift algorithm.Then tracking face with CamShift algorithm adopts the skin color model.If tracking is disturbed or lost target,relocate the target area automatically.This method can improve the robustness and stability of face tracking in video.In terms of single frame face image reconstruction,interpolation reconstruction method and “Hallucinating Face” reconstruction method are researched and implemented.In the“Hallucinating Face” method,some improvements based Baker's method are proposed.In the training process,sampling is improved for 33? overlapping sampling,in the restoration process,and restoration is improved for the image block weighted fusion and reconstructed with layered filtering.The simulation results show that the PSNR of the improved method is2.246 dB higher than the original method,which has better visual effect.In terms of multi frame image reconstruction,the sequences of face images reconstruction based Projection on Convex Set(POCS)algorithm with Speeded-Up Robust Features(SURF)image registration method is researched and implemented.Images are matched by facial character points,and motion is estimated by perspective transform,then the face image is reconstructed.The simulation results show that the method in this thesis can overcome ghosting caused by large scale motion,which can avoid distortion of reconstructed image.In terms the software design and implementation,after analyzing the functional requirement,the software is designed and implemented based on Visual Studio 2010 platformby C++ language and OpenCV visual library,using Qt Creator to design the interface.The face detection and reconstruction in video software including 6 modules,which are the “logging”,the “image operations”,the “video face detection and tracking”,the “single frame face reconstruction”,the “multi frame face reconstruction” and “image quality evaluation”.The software can realize limit user permissions,face image tracking,and face image reconstruction.