Deep learning face representation by fixed erasing in facial landmarks

Face verification (FV) is a challenging problem, because occlusion, posture, illumination, aging will affect the accuracy of FV. Deep convolutional neural networks (DCNNs) have been widely used in many computer vision tasks. Due to the strong feature learning ability, DCNNs improve the accuracy of FV while they also bring some problems such as overfitting. Erasing has been proven to be an effective method for reducing overfitting, while the erasing position is seldom considered. We analyse the effect of different erasing positions and propose a novel data augmentation method especially for FV, called Fixed Erasing(FE). We randomly erase some face images with random size of rectangles centered on fixed facial landmarks such as the centers of eyes, tip of noses and the corners of mouths. Our method can alleviate the risk of overfitting and make the models learn more robust features. And our method can be easily merged with most DCNN-based FV models through a few lines of code. Extensive experiments demonstrate that FE can improve the performance of most recently proposed FV models on several popular benchmarks.