AMLTPCNN: An Adaptive Multi-Scale Local Ternary Pattern-Based Deep Fusion Framework for Illumination Invariant Texture Classification

Abstract

Texture images are inherently affected by lighting and scale variations, which often cause traditional machine learning and state-of-the-art deep learning approaches to fail. We propose a hybrid approach for illumination-invariant texture classification. Our proposed framework combines an Adaptive Multiscale Local Ternary Pattern (AMLTP) with a Convolutional Neural Network (CNN) based texture descriptor. The former learns highly discriminative local features that are, to some extent, illumination-invariant and can encode texture at multiple scales. These are then fed into a CNN to learn a hierarchy of abstract features that provide robustness to illumination changes. The benefits of combining CNNs with handcrafted LTP variants have been shown to retain the local description capabilities of LTP variants and to enable hierarchical feature learning in CNNs, while circumventing the problems individual techniques face. The proposed methodology was evaluated using four well-known datasets: KTH-TIPS2b, CUReT, Outex, and Kylberg. The experimental results demonstrate that AMLTPCNN achieves enhanced classification accuracies of 96.2%, 94.8%, 97.0%, and 98.1%, respectively. This signifies a significant improvement over traditional approaches, including LBP, LTP, CNN, and previously employed hybrid methods. Moreover, this method shows better resistance to changes in lighting and background noise, as demonstrated by the improved precision, recall, and F1-score. The findings corroborate the efficacy of the proposed AMLTPCNN framework, thereby illustrating its capacity to furnish a efficient solution within practical contexts.