A Comparison Study of Image Data Augmentation Techniques in Computer Vision

Image data augmentation is a crucial technique in the field of computer vision. Due to the difficulty of acquiring image datasets in certain domains, this technology holds significant importance in industries such as medicine and manufacturing. Image data augmentation enhances datasets by applying various transformations and processing techniques to original images, thereby improving the model’s generalization ability and enhancing performance in tasks like image processing. This study investigates the impact of image data augmentation techniques on image classification tasks, using the ResNet18 model. Three augmentation methods were evaluated: no augmentation (control), horizontal flipping (traditional method), and Mixup (modern method). The experimental dataset consists of images from five categories: vinegar, sunflower, daffodil, tulip, and orchid, with a total of 4,170 images. The experiments were conducted with 50 training epochs for each method, and the results show the following validation accuracies: $61 \%$ for the baseline (no augmentation), $77 \%$ for horizontal flipping, and $89 \%$ for Mixup. These results clearly demonstrate that modern data augmentation methods provide substantial improvements in classification accuracy compared to traditional methods, while also reducing the risk of overfitting by expanding the dataset. Additionally, recent research progress is introduced, and future developments in image data augmentation are discussed.