Assessing Electroluminescence Image Quality with Machine-Learning and Grey-Level Co-Occurrence Matrix Texture Descriptors

Electroluminescence (EL) based fault detection and power prediction is highly depending on image quality. When it comes to daylight imaging and moving acquisition, however, existing Signal-to-Noise Ratio (SNR) metrics are unreliable; a substantial obstacle for plant-scale automation of postprocessing. This work involves using Grey-Level Co-occurrence Matrix (GLCM) texture features for estimating noise and blur severity in EL images of Photovoltaic (PV) modules. The aim is to directly gauge, rather than predict, image quality. Given the absence of a reliable baseline metric, blur and noise were added artificially to lab quality images, generating exactly known labels for training and validation. Candidate models were kept simple and suitable for linear regression. Hyper-parameters have a great impact on GLCM features and were optimized concurrently with the regression models. To stress-test the model, they were validated against synthesized datasets from completely different and intentionally damaged modules. The proposed method shows a lot of promise, being capable of linearly tracking the noise in the synthesized datasets. For blur, however, it is less accurate. More work is required, including diversifying the applied training models and texture features, and finding a solution to validating them on daylight-captured datasets.