An Approach Of Face Detection Based On Convolutional Neural Network And Conditional Random Fields

Face detection is a complicated pattern detection problem, the main difficulty is caused by different imaging angle, such as in-plane rotation, out-of-plane rotation, it will directly affect the accuracy of the determination face. Current detection method based on deep learning neural network has higher detection rate, but the processing in the output layer of the neural network is inaccurate, it ignores the relationship between the multiple windows of one face,thus makes the final face frame not accurate. This paper is based on the conditional random field models(CRF) and to adjust the output layer of the network, makes the final face frame more accurately.An approach based on convolutional neural network and conditional random field models to detect faces in images is presented in this paper(CRF-CNN).The proposed approach has improved the accuracy of the final face frame. The method firstly trains the convolutional neural network to get a classifier of face and non face, then detects faces from the input images and gets the window which contains face;Secondly, marks all the detection windows of the same face,and take the confidence score of the windows as the random variables of CRF,then uses CRF models to calculate the relationship between windows, makes sure which window should leave behind according to the closeness of the relationship,finally, according to the size of the overlapping area and horizontal distance, vertical distance to merge all the windows of the same scale or different scale of the same face, then gets the final face frame. In order to improve the detection rate, the method scales the input pictures with different scale, different scaling degree only in a small extent affects the detection time, does not affect the correctness of the detection, so the method is not sensitive to choose what kind of scaling algorithm and its parameters.We compared our proposed detector method with the convolutional neural network method for detection of DDFD, R-CNN and deformable parts feature detection method DPM. The experimental results show that the accuracy rate and recall rate of CRF-CNN is similar to DDFD, higher than R-CNN and DPM. In the face detection of in-plane rotation and out-of-plane rotation,the final face frame of CRF-CNN is more accurate, especially in face detection of out-of-plane, the confidence score of DDFD is 0.99232,CRF-CNN is 0.99759, higher than that of DDFD 0.00527 on average.