Analysis of SVD-Based Image Compression for Efficient Deep Learning in Face Mask Classification

Abstract

This study investigates the use of Singular Value Decomposition (SVD) as an image compression technique to improve the efficiency of deep learning models for face mask detection. The proposed approach applies SVD-based compression with different values of k (k = 10, 30, 50) prior to training a MobileNetV2 model. Experimental results show that SVD-based compression can significantly reduce computational cost while maintaining high classification performance. The model trained with k = 50 achieves the highest accuracy of 99.43%, slightly outperforming the model trained on the original dataset. In addition, compressed datasets require less training time, with the fastest configuration (k = 30) achieving a substantial reduction in training duration. The results also indicate that moderate compression levels provide an optimal balance between efficiency and accuracy, while excessive compression leads to performance degradation due to loss of important image features. Furthermore, the training process demonstrates faster convergence for compressed datasets, indicating improved learning efficiency. Overall, this study confirms that SVD-based image compression is an effective preprocessing technique for improving deep learning efficiency without significantly compromising classification accuracy.