A YOLO-Based Local-Safety Branch Framework for Battery Detection in X-ray Images of Small WEEE Devices

Abstract

Preprint. This manuscript has been submitted to the Journal of Cleaner Production and is currently under peer review.Hidden batteries in small waste electrical and electronic equipment (WEEE) are a safety risk because RGB images can identify the device type but cannot confirm whether a battery remains inside the casing. This study evaluates a two-stage safety gate using paired RGB and high-quality (HQ) X-ray images from XBAT+. Stage I uses RGB predictions to determine whether an item belongs to a predefined set of high-prevalence battery-risk product categories. Items predicted as one of these direct-diversion categories are routed to manual sorting or dismantling, where a battery is removed if present. Items for which no direct-diversion category is predicted are not treated as battery-free. Instead, their paired HQ X-ray images proceed to Stage II for internal battery screening using YOLO26 detector-only, object-crop, classifier-assisted fusion, and P2 local-safety branches. Image-level false negatives are the primary safety metric because a missed battery can reach crushing or shredding and trigger a fire that endangers workers, damages equipment, and forces a costly operational shutdown; a false positive instead routes an additional item to manual review. Across YOLO26n/s/m/l/x scales, the detector fused with a crop-trained classifier applied to the full X-ray image reached zero observed false negatives for four of five scales at the default 0.25 confidence threshold. In post-hoc threshold-sensitivity analysis, this branch reached zero observed false negatives for all scales, but with many false positives. The P2 local-safety full-image and object-crop architecture reached zero observed false negatives at the default threshold for all scales, also with high manual-review workload. Overall, zero observed battery misses are achievable in this XBAT+ setting, but only by accepting increased manual review workload.