علم تصمیم گیری و سیستم های هوشمند 3060-7574 2 مجموعه مقالات انگلیسی 2026 02 20 Development of an Intelligent System Based on Deep Learning for Automatic Detection of Fruit Defects Development of an Intelligent System Based on Deep Learning for Automatic Detection of Fruit Defects 1 14 EN Vahid Kameli PhD Student, Department of Electronic Engineering, Faculty of Electrical Engineering, Shahrood University of Technology, Shahrood, Iran Hadi Grailu Assistant Professor, Department of Electronic Engineering, Faculty of Electrical Engineering, Shahrood University of Technology, Shahrood, Ira Journal Article 2025 12 21 <p>The present study introduces a comprehensive system for automatic fruit defect detection using a hybrid architecture that integrates the ResNet50 deep network with CBAM attention mechanisms. ResNet50 serves as the backbone with its 50 layer, 16 block residual structure, while CBAM enhances feature extraction relevant to fruit defects. The study evaluates the impact of CBAM on ResNet50’s performance, demonstrating that the combined model achieves 94.8% accuracy—an improvement of 4.4% over the baseline’s 90.4%. Similar gains appear across other metrics: the F1 Score rises from 89.7% to 94.5%, and sensitivity from 88.2% to 93.6%. Analyses show that channel and spatial attention independently improve accuracy by 1.7% and 1.4%, respectively, but their integration within CBAM produces a synergistic effect accounting for the full 4.4% gain. Overall, the results confirm that attention mechanisms can significantly enhance defect detection in deep neural networks, and that the proposed model offers both high accuracy and strong generalizability to other agricultural products.</p> <p>The present study introduces a comprehensive system for automatic fruit defect detection using a hybrid architecture that integrates the ResNet50 deep network with CBAM attention mechanisms. ResNet50 serves as the backbone with its 50 layer, 16 block residual structure, while CBAM enhances feature extraction relevant to fruit defects. The study evaluates the impact of CBAM on ResNet50’s performance, demonstrating that the combined model achieves 94.8% accuracy—an improvement of 4.4% over the baseline’s 90.4%. Similar gains appear across other metrics: the F1 Score rises from 89.7% to 94.5%, and sensitivity from 88.2% to 93.6%. Analyses show that channel and spatial attention independently improve accuracy by 1.7% and 1.4%, respectively, but their integration within CBAM produces a synergistic effect accounting for the full 4.4% gain. Overall, the results confirm that attention mechanisms can significantly enhance defect detection in deep neural networks, and that the proposed model offers both high accuracy and strong generalizability to other agricultural products.</p>