Noise Invariant Texture Classification Using the Bi-Quad Gabor Filter Hybrid Descriptor

Abstract

Texture classification is an important computer vision technique used in material identification, medical imaging, and quality analysis. But traditional techniques often fail in noise robustness and are incapable of handling complex as well as macro-level texture information. This paper proposes a new approach that combines noise-robust Biquad Gabor filters, a customized CNN, and an SVM to improve the effectiveness of texture classification, especially in the presence of heavy noise. The main contributions of this paper are the combination of advanced techniques such as noise pre-processing, Biquad Gabor filters, a simplified and customised CNN for texture, and multiclass classification using an SVM. The proposed approach improves the reliability of feature extraction and provides noise immunity performance relative to existing techniques. The efficiency of the approach has been proven through other methods. The framework's efficiency has been demonstrated through rigorous evaluations on benchmark datasets such as KTH-TIPS2b, KylbeSintom, Brodatz and CuReT. It achieved accuracies of 95.8%, 94.1%, 97.2% and 93.5% in noise-free environments, while maintaining accuracies of 92%, 90%, 93.6% and 89.2%, respectively, in noisy settings. This comprehensive approach is more precise and expedient than the leading approaches from 2020 to 2024. It is also versatile for multi-class problems and scalable for real-world applications, demonstrating its practical utility. This renders it a robust and effective method for classifying textures in diverse challenging scenarios.