The Research Of Facial Expression Recognition Based On GPU High Performance Computing

In the field of digital image processing, with the development of research content and algorithm complexity, the time and resources consume more, this has brought great challenges for the future research in this field. With the high speed development of graphics processing unit (GPU), it provide a good platform for graphics processing general computing other than image processing, Unified computing equipment structure (CUDA) published by NVIDIA can efficiently make use of GPU strong processing power and huge memory bandwidth graphics to the calculation other than graphics rendering, it is widely used in various fields of modern science and technology. On the other hand, face recognition is one of new research topics about the pattern recognition and image processing. Especially, in this field of face expression recognition, because the expression is very complex, computer identification is not easy, then we used a statistical method to identify face expressions, the data is intensive, it has a large amount of calculation, the repeatability is high, this is the typical characteristics of parallel computing, so we study facial expressions recognition algorithms and optimization method on the GPU performance platform.Firstly we analyze the GPU structure and relevant theoretical knowledge about CUDA, then research and analyze the parallelism of digital images, the experiments of image binary algorithm based on GPU prove that GPU has a significant advantage in digital image parallelization research. We design a high performance computing solutions for face expressions recognition using the general GPU calculation method and CUDA architecture, and realize the face expressions recognition parallelization based on the GPU GTX 260+with 216 stream processor, The method does not affect the precision of original CPU algorithm in the meanwhile, the efficiency of parallel implementation by using GPU can improve 220 times, the experimental results show that the high performance computing based on GPU processing is very effective in the face expression recognition, and can improve significantly the efficiency of calculation. we research the shared memory and texture memory in the GPU to synchronize the thread block. The texture memory storage capacity are large, They can fully meet the requirements of face data stored to a global memory, and compared to the global memory, the delay between threads in the shared memory is only 1/100, in the visit between threads is very fast, so making full use of the advantage local storage resources is a way to achieve minimum transmission delay between threads. this method can optimize GPU algorithm to solve the delay between threads in the GPU, the efficiency have been significantly improved, as compared to the CPU, the speed has nearly 700 times faster, the results of the study show that the GPU have strong adaptability for large-scale data parallel computing, provide a new way to improve the efficiency of pattern recognition.