CartesiaNet: A Lightweight Neural Network for PV-Module Corner Detection

Diagnostic inspection of PV plants using Electroluminescence (EL) imaging for commissioning or selling of PV assets is becoming increasingly widespread, as it has the potential to not only detect physical causes of power loss in a string, be it cracks, finger faults, or material degradation, but also to spot problems before they become severe, and track their progression. However, it is currently costly, requiring both manual imaging of PV modules and inspection of post-analyses. Automation could drastically reduce cost and manual labor time, and increase utility. This research addresses some of the most important automation challenges, by developing a lightweight and fast Convolutional Neural Network (CNN), dubbed CartesiaNet, to locate the outer corners of PV modules in EL images acquired in daylight. This will allow immediate rectification and cropping using quadrilateral transforms. To make it viable for embedded processing on an automated platform, as well as to reduce costly computational load in the cloud, a number of strategies are employed to drastically reduce both network size and number of floating-point operations (FLOPS) while increasing accuracy: Specializing the CNN to target PV module types that share similar features. Synthesizing a varied dataset from a small number of pictures of the target PV modules, by constructing simulated 3D-environments and adding effects like noise, blur and contrast. Using cartesian coordinates instead of segmentation maps as ground truth, to reduce network size and improve convergence, hence the name “CartesiaNet”. Using masking to create additional context outside the region of interest (RoI). The current proof-of-concept network was trained on a 11,264-sample image dataset, synthesized from 4 cropped PV modules of the same type, and shows highly promising results. It detects 97.3% of all corners in the validation set, with an overall accuracy of 0.970%, F1 score of 0.927, mean average precision (mAP) of 0.934 (0.951 with the trapezoidal assumption), and average location error of 5.8 pixels. These scores are much higher than more generic object recognition architectures and could dramatically reduce the need for manual correction, while the number of trainable parameters is less than 1 mio, more than 1 order of magnitude smaller than most modern computer vision architectures. While the main downside of CartesiaNet is, that it is, by design, very specific to certain module types, this is compensated by the extremely high accuracy, ease of training, and the low amount of initial data required.