JoWUA

Abstract

The research work introduces a lightweight Orthogonal Convolutional Neural Network (OCNN) for stationary texture recognition and segmentation through the fusion of GLCM-based statistical feature extraction with orthogonal CNNs. The main aim is to improve the accuracy of texture recognition while at the same time lowering the computational cost as compared to traditional deep learning techniques. This paper focuses on addressing the drawbacks of traditional CNN techniques in capturing stationary texture features as well as feature redundancy. The research is made up of image pre-processing, GLCM feature extraction, orthogonal convolutional feature enhancement, and seven-layer CNN classification. The experiments were performed on the Outex datasets, and the accuracy and robustness measures used include noise, rotation, and illumination changes. Experiment results show that OCNN has an accuracy of 98.0% while AlexNet, VGG, GoogleNet, and ResNet have accuracies of 89.4%, 91.2%, 92.1%, and 93.5%, respectively. On average, the model yields improvements from 4.5% to 8.6% compared with the baseline approaches. Also, statistical validation proves that the method offers consistent results due to 98.0% ± 0.42 of the accuracy and [97.62%, 98.38%] of the 95% confidence interval. Furthermore, robustness evaluation reveals insignificant losses from 1.8% to 2.5% even under unfavorable conditions, which is indicative of excellent generalization ability. In summary, the proposed OCNN model integrates handcrafted statistical features and deep learning together and hence achieves efficiency, non-redundancy, and computational effectiveness.