Segmentation and XAI approach for detecting e-waste from wet biodegradable waste: code, trained weights and results

This deposit accompanies the article "Segmentation and XAI approach for detecting e-waste from wet biodegradable waste", submitted to Waste Management. It contains the code, trained weights and results behind every number reported in the article. The study trains detectors to flag electronic waste in wet organic waste using only synthetic images, made by compositing 54 screened e-waste cut-outs into real organic waste photographs. The detectors are then evaluated on 387 withheld real e-waste photographs and 746 organic photographs. YOLOv11s flagged 84.5% of e-waste photographs at a calibrated 10% false-alarm rate. An ablation over eleven detector configurations and a weighted-box-fusion ensemble showed that adding CBAM attention lowered detection by 11.9 to 16.3 points, and a bidirectional feature pyramid never helped. Contents:- code/: the full pipeline (split construction, cut-out extraction and screening, synthetic dataset generation, training, evaluation, ensembling, latency), model configurations, split manifests and the scripts that produce every figure and table.- weights/: trained weights (best.pt) for all eleven detectors, with training arguments, per-epoch logs and synthetic validation summaries.- evaluation/: per-image scores on the withheld real photographs, threshold sweeps and summaries for each detector and the ensemble.- results/: summary tables across all detectors and latency measurements.- paper_statistics/: statistics quoted in the article, split-conformal calibration and the HiResCAM localisation check.- logs/: cut-out extraction and screening logs, dataset configuration and per-object visible fractions. No images are included. The source photographs come from RealWaste (CC BY 4.0, https://archive.ics.uci.edu/dataset/908/realwaste) and TrashBox (https://github.com/nikhilvenkatkumsetty/TrashBox), which has no stated licence. The cut-outs and synthetic images are regenerated by the seeded pipeline from the split manifests in code/splits/, so all results can be reproduced once both collections are obtained. See README.md for details and code/README.md for run instructions.