Research And Implementation Of Face Detection Algorithm Based On Deep Learning

In rencent two decades,face detection technology has been a hot topic in the field of computer vision.Benefit from Viola and Jones invented the cascade detector,making face detection in a simple environment with real-time and feasibility.However,with the development of society,face detection technology is more and more applied to complex scenes,such as: video stream monitoring,intelligent police car,etc.This complex background on the face detection technology takes high demand in precision and instantaneity.Conventional cascade detectors are no longer meeting the needs of these scenarios.In order to solve the face detection requirements in this complex scene,more and more researchers have applied the convolution neural network to the face detection algorithm.Although these algorithms have achieved fruitful results,the existing face detection algorithm based on convolution neural network still has some problems:1.Still use the traditional sliding window paradigm for the generation of face candidate box,which makes time consumption too large.2.The algorithm used to filter the face candidate box is not efficient enough,which further exacerbates the consumption of time.3.Convolution neural network depth is not deep enough,making the algorithm generalization ability is not strong enough,too deep network makes the time consumption is too large.In this thesis,on the basis of deep convolution neural network,a high performance face detection algorithm is implemented based on GPU technology and cascading idea in traditional detection algorithm.The algorithm can deal with complex cases(light difference,Occlusion,angle changes,etc.)in real time under the face detection.The main contributions of the algorithm are summarized as follows:1.The algorithm uses the GPU technology,and achieves parallel computing module of the candidate box filter algorithm in the GPU.Which greatly speeds up the candidate box filter.2.The algorithm defines a candidate box generating network,generating the candidate box of face in the convolution neural network.It reduces the number of negative sample candidate box generation,and speeds up the generation of the candidate box at the same time.3.Producing a new face classification network by improving the existing convolution neural network,which significantly reduces the time consumption,and only loses a little classification accuracy.4.The multi-task combination is introduced in the network of the algorithm,that is,the classification task is combined with the regression task,which improves the precision of the algorithm.5.Combining the candidate box generating network with face classification network to achieve an end-to-end cascade network,which further enhances the speed of the algorithm.Finally,based on the actual scene requirements,a distributed face detection system based on the algorithm is implemented.It is shown that the distributed face detection system based on this algorithm has high precision,real-time,high concurrency and scalability.