Update what I think are the latest weights are attached here: Segmentation Weights, Classification Weights. Apologies if those are not correct
I also just uploaded the paper to this repository. It is not anywhere near the level of rigor required for a scientific journal, it is really just a 7th grade lab report. Up here for storage and if anyone is interested.
This repository is for the SoilingNet Project. Soiling Net is an AI model to analyze soiling and power loss on photovoltaic panels, with the ultimate goal of making solar panel maintenance easier for everyone.
Credit to the DeepSolarEye project for the dataset we use for this project:
S. Mehta, A. P. Azad, S. A. Chemmengath, V. Raykar and S. Kalyanaraman, DeepSolarEye: Power Loss Prediction and Weakly Supervised Soiling Localization via Fully Convolutional Networks for Solar Panels, 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, NV, 2018, pp. 333-342.
I will add the paper for this project once I have posted it to ArXiv.
We also introduced a subset of the DeepSolarEye dataset with hand labeled semantic segmentation masks for this project (additional samples are generated with data augmentation).
Segmentation model implementations are from the keras-segmentation-library
SoilingNet consists of 2 sub-systems, the first is a semantic segmentation model, which we are able to train in a fully supervised manner, that predicts soiling type and distribution from images of a solar panel. The second system is a classification model that produces a prediction for soiling impact severity.
