Scattering Representation and Attention-Based Residual Learning for Image Classification

This paper introduces the concept of combining wavelet scattering with attention mechanisms in ResNet for improved image feature extraction. The new proposed architecture seamlessly integrates translation-invariant scattering module with an attention-based residual neural network, which is very useful for supervised learning tasks. The framework utilizes scattering priors, enhancing network stability against deformations and preserving high-frequency detail preservation for effective object classification. Attention Residual Learning blocks within the network merge attention and residual learning to create interpretable and discriminative representations. The experimental results on subsets of CIFAR10 dataset validate the effectiveness of proposed work, especially in handling small datasets. This paper highlights the potential of proposed network in advancing image classification, particularly in limited-data scenarios.