<p>Apple variety identification plays a crucial role in pomology and agricultural sciences, as it could effectively assist growers in optimizing orchard management, enhancing product quality, and meeting consumer demand. Traditional identification methods based on visual observation are often influenced by various factors, including human subjective judgment and inter-cultivar variability. To address these challenges, with the support of the China Agriculture Research Systems for Apple Industry and Jiangsu University, we collected sample images of eleven common apple varieties in China, followed by image enhancement and dataset expansion to establish an apple sample database. Subsequently, Convolutional Neural Network (CNN), MobileNet Version 2 (MobileNetV2), and Visual Geometry Group 19 (VGG19) neural network models were utilized for apple variety classification using image-based data. Additionally, two optimization techniques, namely Multi-Head Attention and Gray-Level Co-occurrence Matrix (GLCM), were incorporated to further improve classification accuracy. Results demonstrated that the baseline CNN achieved an accuracy of 96.46%, while MobileNetV2 and VGG19 reached 97.78% and 97.25%, respectively. Multi-Head Attention improved feature extraction but sometimes reduced performance, as observed in MobileNetV2 (87.33%). In contrast, GLCM significantly improved model accuracy, with MobileNetV2 achieving the highest accuracy (98.25%) and the lowest Mean Absolute Error (MAE) (0.0571). GLCM consistently outperformed other techniques across all models, proving particularly effective for texture-rich image classification. These findings suggest that GLCM is a powerful enhancement for deep learning models, improving accuracy, precision, and recall in apple variety classification, with MobileNetV2 combined with GLCM yielding the best overall results.</p>