Research On Multi-view Face Detection Algorithm Under Unconstrained

As an important part of intelligent human-computer interaction technology,face detection is a hot topic in the field of pattern recognition,artificial intelligence and computer vision research,and has attracted more and more researchers' attentions.Poses,as one of the most prominent problems in face detection,introduces nonlinear factors into face detection system,which results in the system performance decreases rapidly.Therefore,in this thesis,we capitalize on Deep Belief Nets to extract face characteristic automatically on the basis of relevance between each part on face,thus avoid too much manual intervention and improve the whole speed of system.Use detector-pyramid architecture instead of traditional softmax algorithm on the classification stage to divide facial pose variation range delicately,which effectively reduce the false alarms in the process of detection.The main works in this thesis is shown as follows:1.Based on the characteristic that PRe LU function has no upper limit,we use PRe LU function which is the improvement of Re LU instead of sigmoid to ease the gradient disappear problem appeared in the training process on traditional deep belief nets when modeling a neuron's output.The additional computation is almost zero and over fitting risk is small.The improvement promotes the network convergence speed and improves the effectiveness of the deep belief network parameters in the training process.Analyze the difference in performance between DBNs optimized by PRe LU and the DBNs activated via sigmoid/Re LU function independently and simulate the experiments from training error rate and convergence.The simulation results show that the PRe LU function optimization scheme performs best.2.Propose a face detection method integrated deep belief nets described in the above paragraph and Float Boost algorithm to solve the problem of missing faces and false alarms caused by poses and occlusion in face detection system.Capitalize on deep belief nets to extract face characteristic automatically on the basis of relevance between each part on face.Float Boost algorithm is used to learn detector of different head rotation.Combine these detectors as a detector-pyramid architecture which adopts the coarse-to-fine and simple-to-complex(top-down in the pyramid)strategy.Input the features extracted by deep belief networks into detector-pyramidarchitecture,then classify whether there is/are face/faces in the image.In order to verify the superiority of the proposed algorithm integrated detector-pyramid architecture,we organize three groups of experiments to analyze the recall and precision in our method and comparisons.The experiment results reveal that the proposed algorithm divides facial pose variation range so delicately that the face in the different rotations can be detected accurately.Moreover,due to the relevance of face feature,not only the poses face is detected,but even when the face partially occluded.The proposed method can be able to achieve similar or better results compared to the other state-of-art methods,without any pose annotations or information about facial landmarks.In above,We know that the proposed algorithm which integrated DBNS and Float Boostal gorithm is affected by unconstrained environment slightly,and has good robustness.