<p>This study introduces a hierarchical deep learning framework for semantic testing of user interface icons in mobile and web applications. In contrast to flat classification approaches, the proposed model leverages a multi-output architecture to simultaneously classify icons by their main categories (e.g., Navigation, Actions) and corresponding subcategories (e.g., search, upload), capturing both structural and functional semantics. To support this task, we present a new benchmark dataset containing 106 icon classes, organized into nine hierarchically structured categories. The model architecture integrates MobileNetV2 for efficient feature extraction, CBAM modules for attention refinement, and Transformer blocks for contextual encoding. Extensive data augmentation and preprocessing are applied to ensure robustness, particularly in class imbalance. Experimental evaluation shows that the proposed framework achieves 96.2% accuracy on main categories, 94.5% on subcategories, and an overall F1-score of 95.8%, surpassing well-established baselines such as ResNet50 and VGG16. Grad-CAM is utilized to visualize class-relevant regions in input icons to improve transparency and interpretability. These visualizations demonstrate that the model’s predictions are grounded in semantically meaningful icon features, supporting its applicability in real-world UI testing scenarios. This work marks the first to offer a complete hierarchical classification solution for semantic icon testing, establishing a new standard in interpretable and domain-specific software quality assurance.</p>