A Multi-Stage Adaptive Copy-Paste Data Augmentation Algorithm Based on Model Training Preferences

Datasets play an important role in the field of object detection. However, the production of the dataset is influenced by objective environment and human subjectivity, resulting in class imbalanced or even long-tailed distribution in the datasets. At present, the optimization methods based on data augmentation still rely on subjective parameter adjustments, which is tedious. In this paper, we propose a multi-stage adaptive Copy-Paste augmentation (MSACP) algorithm. This algorithm divides model training into multiple training stages, each stage forming unique training preferences for that stage. Based on these training preferences, the class information of the training set is adaptively adjusted, which not only alleviates the problem of class imbalance in training, but also expands different sample sizes for categories with insufficient information at different training stages. Experimental verification of the traffic sign dataset Tsinghua–Tencent 100K (TT100K) was carried out and showed that the proposed method not only can improve the class imbalance in the dataset, but can also improve the detection performance of models. By using MSACP to transplant the trained optimal weights to an embedded platform, and combining YOLOv3-tiny, the model’s accuracy in detecting traffic signs in autonomous driving scenarios was improved, verifying the effectiveness of the MSACP algorithm in practical applications.