Improving Unbalanced Security X-Ray Image Classification Using VGG16 and AlexNet with Z-Score Normalization and Augmentation

Addressing the challenge of unbalanced data sets in convolutional neural network (CNN) models for image recognition, this study aims to investigate the impact of data augmentation and normalization. The problem lies in the limited generalization of the model due to class-based data differences, hence the need for several techniques in data preprocessing such as data augmentation and normalization. The main contribution of this research is a comprehensive analysis of the effectiveness of data augmentation and normalization techniques in improving model performance. The research utilized the AlexNet and VGG16 architectures and conducted extensive experiments on data sets with varying degrees of imbalance. Data augmentation generates additional examples, while normalization reduces convergence issues. The results show that training the AlexNet model without augmentation results in a low accuracy of 0.24, underscoring the challenges posed by skewed data distributions. In contrast, augmented data substantially improves performance, with AlexNet achieving an accuracy of 0.91 and VGG16 achieving 0.84. In addition, normalized data also made a positive contribution, showing an accuracy of 0.74 for AlexNet and 0.67 for VGG16. In conclusion, data augmentation and normalization techniques are essential in reducing the effects of data imbalance, thereby improving the generalizability of the models. The improved accuracy of the data using normalization techniques indicates the ability of the model to read the data after normalization. This study underscores the importance of preprocessing strategies in optimizing model performance and advancing the field of deep learning in image recognition.