With the growing popularity and decreasing cost of solar power, crystalline solar panels have been widely adopted in residential and commercial applications. Increased production and prolonged usage of photovoltaic (PV) modules necessitate automatic detection of defects in utility-scale solar power plants. Micro-cracks in particular is are a type of defect that degrade the performance of the modules. This study aims to extend the industrial application of image classification by implementing state-of-the-art convolutional neural network (CNN) architectures and an ensemble of CNNs for identifying micro-cracks from electroluminescence (EL) images of PV modules. Transfer learning has become increasingly popular for mitigating the prerequisite of large training datasets and for performing satisfactorily on smaller, more practical datasets. In this study, pre-trained models like VGG-16, VGG-19, Inception-v3, Inception-ResNet50-v2, ResNet50-v2, and Xception are individually assessed before aggregating them using the ensemble method. Ensemble learning further increases the accuracy while reducing the risk of relying on a single model. The highest accuracies of 96.97% and 97.06% were achieved through the ensemble method for both monocrystalline and polycrystalline solar panels respectively. The individual algorithms have also shown highly accurate performance that is feasible for detecting micro-cracks on PV cells with lower computational cost.