AN ADAPTIVE MULTI-LEVEL DATA AUGMENTATION APPROACH BASED ON MOSAIC, COPY-PASTE, AND MIXUP FOR OBJECT DETECTION

This article addresses the topical issue of data augmentation for object detection systems based on YOLO-family neural networks; specifically, the mechanisms, advantages, and limitations of the Mosaic and MixUp methods are analyzed, along with the practice of their joint application in modern detector implementations. It is established that both methods operate only at the level of recomposing existing training data (scene-level and pixel-level, respectively) and do not add new object instances, which limits their effectiveness for rare and small object classes. An adaptive multi-level data augmentation pipeline is proposed, which complements Mosaic and MixUp with a third, instance-level stage based on the Copy-Paste method, with an emphasis on rare and small object classes. A unified control mechanism has been developed, which dynamically adjusts the probability and intensity of each of the three augmentation levels based on class-frequency and object-size statistics of the current training batch; in particular, formulas for class-balanced instance sampling for Copy-Paste, an adaptive mixing coefficient for MixUp, and a gradual fade-out of the Mosaic probability are proposed. A computational experiment using the YOLOv8n detector on a subset of a dataset with densely arranged small objects was conducted, showing that the proposed pipeline outperforms the Mosaic combined with MixUp configuration without adaptive control by 5.7% in mAP50, by 6.4% in mAP50-95, and by 26.2% in small-object detection accuracy (APsmall), at an acceptable training-time increase of 13%.