Research And Implementation Of Face Detection Based On Convolutional Neural Network

Nowadays,as one of the important methods to implement computer vision,convolutional neural networks have developed vigorously and rapidly.The object detection technology based on convolutional neural network far exceeds the traditional algorithm in terms of accuracy,and can reach or exceed the level of the human eye.As a core technology in the field of object detection,face detection technology is an important part of computer vision,which can be applied to many scenes such as attendance system,intelligent monitoring,station security and so on.As a classic representative of the object detection algorithm based on convolutional neural network,the Faster-R-CNN algorithm achieves high precision in the object detection task,but in the face detection task,this algorithm faces many difficulties and challenges.Especially when there is a small face in the picture,the detection accuracy is very low.In response to this problem,this paper improves the Faster-R-CNN algorithm in the following aspects:(1)Replace the feature extraction network in the Faster-R-CNN algorithm with the residual network structure,so that the overall network can fully extract the features of the face and combine the Feature Pyramid Network with the feature extraction network,which can realize the fusion of the underlying features and the top-level features,which is conducive to the detection of small faces;(2)Aiming at the problem that the small face can't match the rectangular box when using the Faster-R-CNN algorithm,the matching mechanism of the rectangular box is set reasonably to improve the detection accuracy of the face.(3)Improve the loss function of the overall network to make the results of the classification and regression tasks more accurate.The Faster-R-CNN algorithm has been improved as described above,and the trained model on the cleaned Wider Face database can obtain the following results:the mean average precision of the simple,medium,and difficult test set of the Wider Face is: 0.936,0.885,0.626;the mean average precision tested on the FDDB data set is 0.962.It can be seen from these two sets of experiments that the improved algorithm can effectively detect small faces and verify the effectiveness of the proposed method.