A HYBRID HANDCRAFTED DEEP FEATURE FUSION FRAMEWORK FOR ROTATION-INVARIANT TEXTURE CLASSIFICATION

Abstract

Texture classification is an essential undertaking for numerous computer vision applications, such as quality control, biomedical image processing, remote sensing, autonomous navigation, precision farming, and defect detection on surfaces. Even though great advancements have been made in texture recognition by employing deep learning techniques, the problem of recognizing textures at arbitrary rotations still remains challenging. Changes in orientation can affect local spatial relationships and decrease discriminative power of classical handcrafted descriptors and CNN models. Classical handcrafted texture features like Local Binary Pattern (LBP), Grey-Level Co-occurrence Matrix (GLCM), and Gabor filter are able to extract useful information but are known to be less resistant to large rotation angles. On the other hand, deep convolutional architectures learn very discriminative features but require an extensive amount of training samples and are not rotation invariant by nature. Therefore, the design of a hybrid framework that combines rotation-invariant handcrafted features with deep semantics is an important issue in research. This paper introduces a hybrid feature fusion approach for rotation-invariant texture classification utilizing Grey-Level Co-occurrence Matrix (GLCM) and Rotation-Invariant Local Binary Patterns (RI-LBP). and deep ResNet-50 features. First, we pre-process texture images by resizing, gray-scaling, normalization, and contrast improvement to increase feature uniformity. Hand-crafted texture features are extracted using statistical analysis and histograms of GLCM and RI-LBP. At the same time, we extract high-level semantic features from a pre-trained ResNet-50 network. These two sets of features are concatenated to a single vector and then normalized to compensate for feature scale differences. A multi-class Support Vector Machine (SVM) with a radial basis function kernel performs the final classification owing to its excellent generalization capability in high-dimensional feature spaces. The proposed framework is evaluated on the benchmark KTH-TIPS2b and Kylberg Sintorn texture datasets, both of which contain significant intra-class variations in scale, illumination, and rotation. Experimental evaluation includes overall classification accuracy, precision, recall, F1-score, confusion matrix analysis, and five-fold cross-validation. The proposed hybrid framework demonstrates superior robustness against rotational variations while maintaining competitive computational efficiency. The combination of handcrafted statistical descriptors with deep convolutional representations enables the model to capture both local structural patterns and global semantic information, thereby improving classification performance compared with individual handcrafted or deep-learning approaches. The proposed methodology provides an effective and computationally efficient solution for rotation-invariant texture analysis in real world imaging applications.