Abstract: The Multi-Task Cascade Convolutional Neural Network (MTCNN) face detection algorithm is often used in some face recognition tasks because it takes into account the detection speed and accuracy. However, in the face of some complex face detection tasks, the real-time performance and accuracy of the network detection still cannot meet the practical requirements. In order to solve this problem, an improved multi-task cascade convolution neural network face detection algorithm is proposed. This method improves the R-Net and O-net modules in MTCNN, and optimizes the NMS algorithm of these two network modules into a better NMS algorithm, that is, the classification confidence of image candidate frames is modified again to avoid the missed detection for face candidate frames with IOUs larger than the preset threshold. The proposed MTCNN face detection algorithm and other comparative algorithms are trained and evaluated on WIDER Face and FDDB datasets. The experimental results show that the improved method can better eliminate the redundant candidate frames and retain the regression window with higher accuracy in the face detection process, which can improve the accuracy of face detection without losing its robustness.