Improving ResNet Model Accuracy with Curriculum Learning Using Blurred Images.

Deep learning has achieved remarkable success in a variety of computer vision tasks, including image classification, object detection, and semantic segmentation. One of the major directions in deep learning development is to design effective training strategies that can improve model performance and generalization. In this paper, we creatively apply curriculum learning using blur images to train the famous ResNet models. We hypothesize that presenting blurry images using curriculum learning can help the model to learn general representation rather than detailed features and to descend the loss closer to the global minimum without overfitting. We verify our hypothesis by training on real-world datasets using control group, and thoroughly examine our discoveries. Our work represents an important step in curriculum learning in the latest ResNet studies, and has the potential to universally improve the performance of image classification models.