Enhancing Thermal Image Classification with Novel Quality Metric-Based Augmentation Techniques

Thermal image classification is critical in various applications, particularly fault detection and monitoring systems such as photovoltaic (PV) modules. However, a common challenge in these fields is the limited availability of large-scale, labeled thermal image datasets. To address this, color image augmentation is widely adopted in machine learning to artificially increase the size and diversity of training datasets, improving model generalization. Traditional augmentation techniques, such as geometric transformations, provide some benefits but may fail to fully capture the unique characteristics inherent in thermal images, which often have lower contrast and different noise patterns than visible spectrum images. So, we argued we need to develop a novel augmentation technique for thermal imaging, where data collection is costly and time-consuming. Our research proposes a novel offline augmentation technique guided by quality metrics to enhance the performance of thermal image binary classification models. By leveraging domain-specific quality metrics, such as image clarity, thermal contrast, and noise levels, we optimize the oversampling process for thermal datasets. For example, starting with a dataset of x images, we generate y additional thermal images, resulting in a total of x + y images used to train the deep learning classification framework. Using a dataset of PV module defects, we demonstrate the effectiveness of our quality metric-based oversampling strategy across several state-of-the-art image classification networks. Our approach outperforms traditional augmentation methods regarding classification accuracy and robustness, including geometric transformations and standard image enhancement techniques. The practical implications of our research are significant, as it provides a more effective and efficient way to improve model performance for thermal imaging tasks, mainly when data availability is limited.