Teacher-Student Mutual Training for Semi-Supervised Object Detection Based on PPYOLOE

With the continuous advancements in deep learning，object-detection technology based on convolutional neural network has become a research hotspot in the field of computer vision. Currently，mainstream object-detection algorithms rely on supervised learning and training models on extensive labeled data. However，unlabeled data are easy to obtain，while labeled data are usually challenging，time-consuming，and labor-intensive to collect. This study proposed a semi-supervised object-detection(PPYOLOE-SSOD)algorithm based on teacher-student mutual training to easily obtain data annotations. First，the student and gradually improved teacher models were trained simultaneously. The teacher model was then used to filter high-quality pseudo labels，which guided students during model training and extracted information from unlabeled images. Further，the exponential average method was used in each iteration to update the teacher model parameters to reduce the instability of parameter transfer. In addition，different data-augmentation methods were introduced to enhance the anti-interference ability of the network. Finally，the unsupervised learning branch was added for the learning of unlabeled data，and the features predicted by the model were processed using an intensive learning method. By sorting the classification features predicted by the teacher model，high-quality features were automatically selected as the pseudo labels generated by the teacher model，thus avoiding the tedious post-processing of pseudo labels and improving the accuracy and training speed of the network. On the MS COCO dataset，the accuracy of the PPYOLOE is improved by 1.4%，1.6%，and 2.1% on 1%，5%，and 10% labeled datasets，respectively，using the semi-supervised learning method. Compared with other SSOD algorithms，PPYOLOE-SSOD achieves the highest accuracy. The source code is at https://github.com/ wjm202/PPYYOLOE-SSOD. © 2024 Beijing Institute of Technology. All rights reserved.