Improved Efficient Convolutional Neural Networks for Complex Scene Mask-Wearing Detection

Abstract

: To solve the problem about low accuracy of mask wear detection under complex lighting and face lean conditions, a method of mask wear detection under intricate environment using efficient convolutional neural network is proposed, which uses pre-training such as hard negative mining to learn more samples of face feature, utilize multi-task convolutional neural networks (MTCNN) to estimate the possibility of face information, and get accurate face location. With attention mechanism in feature pyramid network, enhanc-ing the weight of key points on human face, employing efficient neural network detection will be wore on mask-wearing detection as a simple binary classification problem. Under the environment of TensorFlow platform, not only data training, data preprocessing, but also the contrast experiment with AIZOO method are completed. A data set containing with 816 images is collected, marked and trained. During the data pre-processing, images are set as fixed size to reduce the amount of computation and promote the detection speed. Then, image enhancement algorithm is used to conduct distortion processing to improve the robust-ness of this model. On this basis, MTCNN is used to detect the face information in pictures, modify and normalize all data, then put them into neural network and the trained model to detection. The experimental results show that under complex conditions such as complex lighting and face tilt, the accuracy can reach 83% and 91% respectively, which means can accurately detect whether wearing a mask.