A phase preserving dual tree complex wavelet convolutional neural network framework for texture classification
S. Gowthaman, Abhishek Das · Results in Engineering · 2026
Wavelet Convolutional Neural Networks (WCNNs) represent a class of models that integrate multiscale wavelet coefficients with convolutional layers channels to improve texture classification. Several hybrid CNNs demonstrated that the incorporation of wavelet decompositions can improve frequency-aware feature extraction. However, existing WCNNs mostly rely on real-valued Discrete Wavelet Transforms (DWT), which capture limited directional information and suffer from shift variance, thereby limiting their ability to represent complex texture patterns with rich directional information. To overcome this limitation, we propose a novel hybrid WCNN method that incorporates the Dual Tree Complex Wavelet Transform (DTCWT) with the Complex-Valued Convolutional Neural Network. Our method overcomes the shortcomings of the WCNN by considering more directional sub-bands from DTCWT and preserving phase information through the CVCNN blocks for the entire network. To validate our proposed method, we evaluate our approach on two standard datasets, KTH-TIPS2b with 11 classes and DTD with 47 classes, where it achieves classification accuracies of 73.22% and 34.3%, respectively. Our DTCWT-CVCNN method achieves near state-of-the-art results accuracy with 0.23 million parameters, which is 97.7% reduction in the parameter count when compared to WCNN.