A hybrid deep learning and ensemble framework for thermal image-based metal detection and classification system

Accurate metal identification and classification is essential for industrial inspection, material sorting, and quality control applications. Conventional metal identification and classification techniques typically rely on laboratory-based analytical methods that require expensive equipment, manual intervention, and significant processing time, thereby limiting their scalability and real-time applicability. To overcome these challenges, this study proposes a novel thermal image–based metal identification and classification framework integrating automated temperature extraction with hybrid deep learning models. Thermal images of different metals were first captured under controlled experimental conditions. Subsequently, temperature values for each metal category were recorded at fixed time intervals using the proposed automated temperature extraction algorithm. The thermal behavior of metals is analyzed by considering their heat conductivity and specific heat characteristics. Based on the collected temperature data, a dedicated and structured dataset was created to facilitate accurate metal classification. For classification, a hybrid architecture combining TabNet representation learning, Autoencoder-based feature extraction, and XGBoost classification was implemented to enhance discriminative capability. Additionally, a hybrid framework combining of Deep Neural Network and Random Forest hybrid model. Experimental results demonstrate that, the two proposed hybrid models achieved high classification accuracy of 99%, outperforming traditional machine learning approaches. The stable convergence behavior and minimal generalization gap further confirm the robustness and reliability of the framework. Overall, the proposed system offers a fast, low-cost, and accurate solution for automated thermal image–based metal identification and classification.