ELPV Detection
4-class defect grading on industrial electroluminescence solar cell imagery. Designed a domain-specific preprocessing transform and combined HOG-SVM with a VGG16 transfer-learning CNN via a stacking ensemble — with separate models trained per crystal type (mono / poly).
4-class defect grading on held-out test set via Random Forest stacking of SVM + CNN outputs.
HOG + SVM (classical CV) and VGG16 CNN (deep learning) fused as an 8-dim probability stacking input.
Domain-specific transform: unsharp masking → background subtraction → histogram-adaptive thresholding.
Approach
The ELPV dataset contains greyscale EL images of solar panels labelled with a defectiveness probability and crystal type (monocrystalline or polycrystalline). The key insight driving the architecture was that these two crystal types produce visually distinct defect signatures, so training a single global model wastes capacity. Instead, separate SVMs and separate CNN fine-tune passes were used for each type, with the outputs merged only at the ensemble stage.
The stacking ensemble concatenates the 4-class probability vectors from each model into an 8-dimensional feature and trains a second-level classifier (Logistic Regression, Random Forest, Gradient Boosting, Bagging were all evaluated) to learn the optimal combination. The DBHTransform preprocessing step was designed specifically for EL imagery: it subtracts a mean-value background and applies histogram-peak thresholding to surface defect contrast that standard normalisation suppresses.