Enhancing object detection and classification with thermal imaging: a deep learning and optimization-based framework

Abstract The proposed research presents a robust and efficient framework for object detection and classification in thermal imaging, leveraging advanced deep learning and optimization techniques. Thermal imaging poses unique challenges, including low-resolution input, varying heat signatures, inefficient feature extraction, and reduced accuracy in real-time performance. Additionally, noise interference, inconsistent thermal scaling, and poor contrast further complicate preprocessing. Traditional methods often fail to capture the nuanced thermal features necessary for accurate analysis. To address these challenges, the proposed methodology introduces a comprehensive preprocessing pipeline specifically tailored for thermal images, incorporating greyscaling, cropping, resizing, noise reduction, image normalization, and contrast enhancement to improve data quality. Improved CornerNet is adopted for effective feature extraction, capturing vital thermal patterns. The Binary Sand Cat Swarm Optimization Algorithm (BSCSOA) ensures optimal feature selection by reducing dimensionality while preserving relevant thermal information. For object detection, the enhanced YOLOv9 network is used, delivering high accuracy and real-time performance. Classification is handled by a Stacked Convolutional Bidirectional Long Short-Term Memory (SCBi-LSTM) network, which effectively captures both spatial and temporal thermal dynamics. Precise object segmentation is achieved through the Enhanced Manta-Ray Foraging Optimization Algorithm (EMRFOA), accurately delineating heat-based object boundaries. This integrated approach significantly improves the accuracy, efficiency, and real-time capability of object detection and classification in thermal imaging, making it suitable for a wide range of critical applications such as surveillance, industrial inspection, and medical diagnostics.