This paper introduces a novel approach to texture classification using wavelet packet signatures. The authors evaluate the performance of wavelet packet spaces in classifying 25 natural textures, focusing on both energy and entropy metrics. They find that an analyzing function with longer FIR (longer filter length) is more efficient for representation and discrimination compared to shorter functions. Energy signatures computed from a minimal set of wavelet packet nodes (17 nodes at four levels) achieve perfect classification, while redundant representations (e.g., using nodes from multiple levels) result in classification errors. The study also demonstrates that a two-layer neural network classifier performs best with three hidden nodes when using the $D_{20}$ analyzing function. The results suggest that wavelet packet representations, particularly with longer analyzing functions, are effective for robust texture classification and subtle discrimination.This paper introduces a novel approach to texture classification using wavelet packet signatures. The authors evaluate the performance of wavelet packet spaces in classifying 25 natural textures, focusing on both energy and entropy metrics. They find that an analyzing function with longer FIR (longer filter length) is more efficient for representation and discrimination compared to shorter functions. Energy signatures computed from a minimal set of wavelet packet nodes (17 nodes at four levels) achieve perfect classification, while redundant representations (e.g., using nodes from multiple levels) result in classification errors. The study also demonstrates that a two-layer neural network classifier performs best with three hidden nodes when using the $D_{20}$ analyzing function. The results suggest that wavelet packet representations, particularly with longer analyzing functions, are effective for robust texture classification and subtle discrimination.