Fig. 8.

Curriculum learning using motion artefacts generated with various levels of severity. (a) The traditional way to train a model fails to consider the complexity of image quality detection where introducing noisy or difficult samples early in training may impair model performance. (b) The training data is divided into different difficulty levels based on a predetermined curriculum. The training procedure progresses from easy to hard image samples, which guides the model to achieve better performance. (The illustration of complexity is shown in Fig. 7).