多任务级联卷积神经网络（multi-task convolutional neural networks, MTCNN）人脸检测算法虽然在一些人脸检测任务上取得了不错的成绩，但是面对一些复杂的小尺度、多人脸检测任务，人脸检测的准确率仍有待提高。提出一种改进的MTCNN算法，首先，在制造数据集时微调交并比（intersection over union, IoU）阈值参数，来更精确地对人脸样本进行分类；其次，用卷积层替换网络的最大池化层，来提升网络性能；最后，在O-Net网络中引入SE（squeeze-excitation）注意力机制，提高网络的特征表达能力。试验结果表明：改进后的算法与原始MTCNN算法相比较，P-Net网络和R-Net网络检测准确率提高了1%，O-Net网络检测准确率提高了0.5%，且改进后的算法在实际人脸检测任务中表现更好。

Although the multi-task convolutional neural networks (MTCNN) face detection algorithm has achieved good results in some face recognition tasks, the accuracy of face detection needs to be improved in the face of some complex small-scale and multi-person face detection tasks. An improved MTCNN algorithm was proposed. Firstly, the intersection over union (IoU) threshold parameter was fine-tuned when creating the data set to classify face samples more accurately. Secondly, replacing the max pooling layer of the network with convolutional layers can improve network performance. Finally, the squeeze-excitation(SE) attention mechanism was introduced into the O-Net network to improve the feature expression ability of the network. The test results show that compared with the original MTCNN algorithm, the detection accuracy of the P-Net network and R-Net network of the improved algorithm has increased by 1%, and the detection accuracy of the O-Net network has increased by 0.5%. Moreover, the improved algorithm performs better in the actual face detection task.