Reliable E-Waste Image Classification via Calibrated Deep Ensembles and Statistical Validation

With such a high rate of E-Waste accumulation, reliable automated sorting systems are imperative. This study proposes a classification framework for e-waste using a hyperparameter-optimised CNN, transfer learning with ResNet50, and uniform probabilistic ensembling. The ensemble achieves higher peak clean test accuracy (95.67%, with a bootstrap accuracy of 95.02% and a 95% CI of [92.33%, 97.33%]) than the individual models on a 10-category dataset of 3000 images with strict partition isolation (0.00% cross-contamination). Pairwise McNemar tests with Holm–Bonferroni correction show that the ensemble outperforms ResNet50 (p=1.20×10−3) and the custom CNN (p=8.33×10−20, where pHolm are Holm–Bonferroni corrected p-values and αadjusted=0.0167 is the adjusted significance level. Predicted probabilities are temperature-scaled to ensure industrial reliability, and an uncertainty-aware rejection filter handles ambiguous or corrupted input with strong robustness to spatial occlusion, blur, and illumination shifts. In addition, the edge deployment metrics are shown to be viable for real-time industrial use: latency for inference is 3.11 ms per image, and throughput is 321.72 images/s. The validation is comprehensive and shows that the calibrated ensemble provides an accurate, robust, and statistically verified solution for automated e-waste recycling, using Expected Calibration Error (ECE), Brier scores, and confidence-gated rejection metrics.