A Shift-Invariant Robust Deep Framework with Improved Classification Performance

This study introduces a hybrid approach by integrating scattering transform into the early layers of the VGG and ResNet neural networks. This advanced method not only addresses translation instability but also enhances feature extraction. Additionally, anti-aliasing technique is included to increase network stability and resilience against input distortions. Through the integration of scattering network, and anti-aliasing technique, the proposed networks demonstrate substantial improvements in accuracy, generalization, and robustness. This underscores the promise of combining traditional signal processing methods with contemporary deep learning architectures for future enhancements in convolutional neural networks. The results from experiments conducted on the CIFAR-10 and Skin Cancer Images dataset confirm the efficacy of the proposed methodology. The study highlights the potential of the developed network to advance image classification methodologies.