Local Binary Patterns (LBPs), while foundational in image processing, face limitations including generalization issues and restricted neighborhood sampling. To address these, we introduce three novel feature extraction techniques: Horizontal Orthogonal Information Fused LBP (HOI-LBP), Vertical Orthogonal Information Fused LBP (VOI-LBP), and Multiplied Orthogonal Information Fused LBP (MOI-LBP). Our methods integrate orthogonal pixel information from the outer 5 \(\times \) 5 neighborhood with the traditional 3 \(\times \) 3 LBP operator through logical bitwise fusion. This approach leverages AND and OR operations on cyclic 8-bit pairs of corner, horizontal, and vertical pixels surrounding the central pixel, significantly enhancing feature richness and noise robustness. Experimental evaluations on benchmark datasets using deep learning (DL) models (VGG-16, MobileNet, EfficientNetB0) demonstrate the efficacy of our techniques. The proposed HOI-LBP, VOI-LBP, and MOI-LBP consistently outperform regular LBP, showing an average increase of 6.63% in training accuracy and 4.7% in validation accuracy. Furthermore, notable reductions in training and validation losses signify improved convergence and generalization capabilities. These results confirm that our fused features provide a more comprehensive representation of image content by capturing subtle variations in texture, gradient energies and structural patterns. Thus, marking a significant advancement in image processing and machine learning (ML) applications.
