Analyzing Impact of Data Augmentation Techniques in Computer Vision

Data Augmentation is a prevalent practice within computer vision, which uses transformations like random flipping, rotation, jittered colors and advanced techniques like Mixup and CutMix to artificially increase a training dataset. This, in turn, seeks to give generalization skills, minimize overfitting, and amplify the diversity of the data. In this paper, we explore the effectiveness of different data augmentation techniques for computer vision applications significantly concerning their effects on image classification tasks using convolutional neural networks (CNNs). The study bases its analysis on Residual Network-18 (ResNet-18), Visual Geometry Group-16 (VGG-16), Visual Geometry Group-19 (VGG-19), Densely Connected Networks-121 (DenseNet-121), and a Modified CNN. These were all trained and evaluated on the Canadian Institute for Advanced Research –10 (CIFAR-10) dataset. According to the results, the modified CNN model consistently attains the highest accuracy of the different tried models. Although its performance record of 90.9% with training augmented and 91.14% without augmentation is better than others, the results show an outstanding generalization capability of the model in different environments indicating resilience and adaptability of the model across an even wider variety.