Data Augmentation on Intra-Oral Images Using Image Manipulation Techniques
Samah AbuSalim, Nordin Zakaria, Norehan Binti Mokhtar, Salama A. Mostafa, Said Jadid Abdulkadir · 2022
The quality, quantity, and relevance of training data determine how well most ML models perform, and deep learning models. One of the most frequent problems in implementing machine learning, though, is a lack of data. This is because gathering such data can frequently be expensive and time-consuming. The diversity of training data for machine learning algorithms is increased through data augmentation without the need for new data collection. Basic image manipulation techniques, including horizontal flip, Brightness and contrast, Noise injection, and histogram equalization techniques were used in this work to produce an augmented intraoral dataset. Faster R-CNN, a CNN-based model, was used to analyze the performance of the data augmentation strategies. An extensive simulation shows that the augmented dataset achieves better accuracy than the original dataset. The experimental results show a mean average precision (mAP) of 72.4 % on augmentation data.