A Lattice-structure-based Trainable Orthogonal Wavelet Unit for Image Classification

This work introduces Orthogonal-LatticeUwU, a novel trainable wavelet unit for enhancing image classification and anomaly detection in convolutional neural networks by reducing information loss during pooling. The unit employs an orthogonal lattice structure, relaxing the zero-at-π condition and decreasing the number of trainable wavelet coefficients. This innovation is a key novelty of the work. The unit modifies convolution, pooling, and down-sampling operations. Implemented in ResNet18, it improved detection accuracy on CIFAR10 (by 2.67%), ImageNet1K (by 1.85%), and the Describable Textures dataset (by 9.52%), showcasing its advantages in detecting detailed features. Similar gains were seen in ResNet34 and ResNet50 implementations. For anomaly detection in hazelnut images on the MVTec Anomaly Detection dataset, the proposed method achieved a segmentation area under the receiver operating characteristic curve of 97.15% and better anomaly localization. The method excels in detecting detailed features, despite increased trainable parameters from using one-layer fully convolutional networks for feature combination.