Résumé: Random erasing uses rectangular areas to erase images, which improves model generalization performance, but there is too much loss of original image information. In this paper, we introduce a novel approach, the improved data augmentation method based on random walk algorithm, called random walk-based erasing (RWE), for training the CNN model. Compared with random erasing, RWE has more information about the original image because it uses scatter points to erase the original image instead of rectangular boxes. We execute experiments related to random erasure and deploy them on existing models, aiming to explore factors that benefit model accuracy. Firstly, we execute a test to determine parameter setting and validate compatibility. Then, the results of our methods on image classification and object detection tasks were obtained. Finally, we compare the impacts that random erasing and RWE make on image classification and object detection tasks, respectively. The results of different tasks indicate that our methods increase the accuracy by ?2.00% on CIFAR-100 and map by ?0.5% on the VOC-2007 dataset.

Langue: Anglais