Efficient Classification of Photovoltaic Module Defects in Infrared Images

Photovoltaic (PV) modules are crucial components in solar power generation, where their efficiency and safety can be compromised by various defects. Infrared images of PV modules can be used to classify these defects, as each type of defect images has different spatial characteristics. This paper presents a novel approach to classifying infrared images of PV modules based on the modern convolutional neural network architecture, ConvNeXt. Many previous studies have initialized models using ImageNet pre-trained weights to deal with PV module datasets containing a small number of training images. However, this requires resizing the images in the PV module dataset (24 × 40) to the size used for ImageNet pre-training (224 × 224), which significantly increases floating point operations for both training and inference. To address the problem, we propose a task-specific pre-training procedure that does not rely on external datasets such as ImageNet, thereby can avoid excessive increase in image size. We also introduce attention mechanisms into the ConvNeXt block to further improve classification performance. Experimental results on the PV module dataset demonstrate that the proposed method achieves competitive performance in classifying multiple types of defects with fewer floating point operations than conventional approaches, offering a computationally efficient and fast solution for PV plant monitoring systems.