Abstract
Aiming at the problem of low accuracy of face detection under complex occlusion conditions, a double-channel occlusion perceptron neural network model was proposed. The area occlusion judgment unit is designed and integrated into the VGG16 network to form an occlusion perceptron neural network. Thereupon, the features of unoccluded regions and less occluded regions in facial images are extracted by the perceptual neural network. Transfer learning algorithm is utilized to pretrain parameters of the convolution layer to reduce the overfitting problem caused by insufficient training data samples. Face features of the whole face were extracted by optimizing the residual network, and then the face features of the occluding perceptron neural network and the residual network were weighted and fused. Experiments were carried out on two open data sets, AR and MAFA. The results demonstrate that the detection accuracy of this method is higher than that of other methods, and the detection speed is faster.
Highlights
As a research hotspot in the field of computer vision, face detection is the basis of face analysis tasks such as face retrieval [1], face alignment [2], human-computer interaction [3], and face superresolution reconstruction [4]
Literature [13] put forward a large margin cosine loss function, which improves the discriminant ability of the traditional Convolutional Neural Network (CNN) network softmax layer. e above methods further improve the performance of CNN face detection, some of which even surpass human recognition ability in some specific data sets
Compared with the other four methods, the detection accuracy of the proposed method is higher, indicating that the detection accuracy of the occluded face can be improved effectively by shielding the face feature element damage caused by local occlusion
Summary
As a research hotspot in the field of computer vision, face detection is the basis of face analysis tasks such as face retrieval [1], face alignment [2], human-computer interaction [3], and face superresolution reconstruction [4]. Literature [12] proposes an orthogonal embedded CNN network, which enhances feature learning of age-invariant deep face. Literature [13] put forward a large margin cosine loss function, which improves the discriminant ability of the traditional CNN network softmax layer. E above methods further improve the performance of CNN face detection, some of which even surpass human recognition ability in some specific data sets. The face detection method based on CNN has achieved great success, the depth feature does not have feature invariance under the influence of light, posture, expression, occlusion, and other factors. With the improvement of occlusion degree, the difficulty of face detection increases, making it difficult for more advanced face detection methods [14, 15] to accurately determine the location of occluded faces
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
