Rotating face detection based on convergent cascaded convolutional neural network

To solve the problem of low accuracy of multi-scale rotating face detection under complex conditions such as large-scale pose change and large-angle face rotation-in-plane, a rotating face detection method based on parallel cascade convolution neural network is proposed. Using a coarse-to-fine cascading strategy, multiple shallow convolutional neural networks are cascaded in parallel on multiple feature layers of the main network SSD. Face/non-face detection, face boundary box position update and face RIP angle estimation are gradually completed. Experimental results on Rotate FDDB dataset and Rotate Sub-WIDER FACE dataset show that the proposed method achieves advanced face detection. The detection precision of the method is 87.1% and the speed is 45 FPS when 100 false positives occur in the rotating Sub-WIDER FACE dataset, which proves that the method can achieve accurate rotating face detection with low time loss.