A lightweight visual perception method for real-time sorting of non-magnetic crushed materials from waste refrigerators

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Non-magnetic crushed materials from waste refrigerators have irregular boundaries, overlapping fragments, reflective surfaces, and an imbalanced class distribution, which complicate instance-level visual recognition. This study develops YOLOv11-MEDSeg, a lightweight instance-segmentation model, and constructs the CMFR dataset containing 800 images and 11,009 annotated instances in six material classes. The model combines C3k2-MDConv2 for adaptive spatial sampling, C2PSA-EMA for feature reweighting, and a Dynamic Progressive Loss Reweighting Strategy (DPLRS) for training. On CMFR, the selected model achieves a mask mAP50-95 of 79.8 %, compared with 76.3 % for YOLOv11n-seg, with reported precision and recall of 97.7 % and 97.6 %, respectively. Across five seeds, mask mAP50-95 is 79.80 ± 0.07 % (mean ± sample standard deviation). The updated FP32 benchmark reports 2.82 million parameters and a mean batch-size-one inference latency of 8.002 ms on an RTX 5070. The model supplies material labels, locations, and instance masks for potential sorting control. Physical sorting purity, mass recovery, and closed-loop throughput remain to be validated.