A Grid-Based Random Erasing Approach for Enhanced Image Classification

In modern image classification tasks, the data is insufficient due to time or money cost constraints, which affects the generalization and accuracy of the model. This paper proposes a random erasure method based on gridded images, aiming to expand the dataset to improve the generalization ability and classification accuracy of the model while reducing the damage to the original structure of the image caused by continuous erasure. This paper uses YOLOv8n-cls, YOLOv8m-cls, and ResNet50 as models and selects CIFAR-10, CIFAR-100, FER2013, and mini-ImageNet as datasets for comparative experiments. The Top-1 accuracy, Top-5 accuracy, and loss function curves on the validation set are used to analyze the effect of random erasing on each model's performance on various datasets. The findings demonstrate that lightweight models greatly increase accuracy on simple datasets like CIFAR-10. However, to effectively use the enhancing effect of random erasing on complicated datasets, more sophisticated models are required. This study shows the impact of random erasing based on gridded images on improving the generalization ability and classification accuracy of the model in image classification tasks. It points out the limitations of this method and potential future optimization directions to better handle more complex classification tasks.