Background: Knee osteoarthritis (KOA) is characterized by cartilage degradation and joint-space narrowing, resulting in increased friction and observable structural damage. Methods: This study introduces a composite hybrid framework for the automatic classification of KOA severity using anteroposterior knee X-ray images. The methodology applies joint-centered cropping and data augmentation to standardize inputs and uses class weighting to mitigate class imbalance. Deep features extracted from fine-tuned ResNet-101 and EfficientNetB7 models are integrated with handcrafted Gray Level Co-occurrence Matrix (GLCM) texture descriptors, and the final predictions are obtained using a soft-voting ensemble. Results: the proposed ensemble achieves 73% test accuracy (macro-F1 ≈ 0.70; weighted-F1 ≈ 0.73) in a four-class setting (KL-0, KL-2, KL-3, and KL-4). Additional experiments across different classification setups demonstrate consistent performance trends, while Grad-CAM indicates that the model primarily focuses on the joint region. Overall, Conclusions: combining ensemble deep learning with complementary handcrafted texture features provides a reliable and interpretable approach for grading radiographic KOA severity.