Face detection is an important domain in computer vision.It has been widely hired in the field of industry,consumers' daily life,and security solutions.Also,the difficulties lying in runtime efficiency and detection accuracy make face detection full of research value.This dissertation made intensive discussion as follows on some difficulties in the practical application of face detection when confronting complex scenes.1.By analyzing and testing various performances among traditional Adaptive Boosting based algorithms and deep learning based algorithms,the cascade light Convolutional Network has been chosen for face detection because of its high balance between detection rate and running time consuming.2.Attempt to increase the detectability of algorithm when facing occluded faces.The Grid Softmax Loss function,Soft Non-Maximum Suppression,and hard example production have been used to promote robustness of the network towards partly masked faces.3.Discussion of difficulties in tiny face detection against common size object detection.A branched cascade Convolutional Network in which the tiny face detect branch has been trained by contextual image information has been designed to solve the challenge of multi-scale friendly face detection along with hard example mining.4.To test performance of the modified algorithm above in practical circumstance,a real-world surveillance video database and a face recognition algorithm has been hired to simulate a real-time face recognition system.The modified face detector has been assembled with a fast object tractor and an image quality evaluator,which consist of the Average Point Acutance score and the Softmax Confidence score,to form a fast face detecting and high quality face screening module,which has been proved to be highly effective when operating in the face recognition system as a front end part. |