Research on Image Quality Assessment Methods for Small Target Data Augmentation under the YOLO Framework
Zhibin Liu, Pengyu Liu, Bifeng Cui, Gang Li, Jun Deng, Qian Li · 2024
When using image scaling algorithms for data augmentation on small targets within the YOLOv framework, it was observed that the ranking of different image scaling algorithms based on their data enhancement effects significantly diverges from the rankings provided by mainstream image quality assessment methods based on the human visual system. This indicates that the machine vision system of YOLOv5 has different requirements for image quality compared to the human visual system. By analyzing various image quality assessment methods, this study identified an image quality assessment method, BLIINDS-II, that is highly correlated with the data enhancement effects of YOLOv5. It was found that YOLOv5's assessment of image quality is derived from the non-DC coefficients of the image following a Discrete Cosine Transform (DCT), focusing on the high-frequency components of the image.