Background Activation Leakage of CNNs and EigenCAM-Guided Weight Decay Control in Amorphous Pork Primal Cut Detection

In pork pre-packaging lines, the recognized cut type determines the packaging material and line branching, but detection is challenged by deformable tissue, dense crowding, a fixed industrial background, and strict latency limits. At the same input resolution and latency, the transformer-based RT-DETR-L localized cuts more accurately than the convolutional EfficientDet-D3, whereas EfficientDet-D3 missed fewer cuts at matched precision for every seed, although not significantly. EfficientDet-D3 also showed spatial activation leakage toward fixed background structures. To suppress it at no inference cost, we propose an EigenCAM-guided two-stage weight decay control protocol: Stage 1 defines a transition zone of the weight decay coefficient λ solely from the validation-set target activation ratio (TAR), and Stage 2 optimizes the training schedule within it; the zone was reproduced across three seeds. At unchanged latency and accuracy (test mAP50 0.984 ± 0.008 vs. 0.978 ± 0.004), the selected configuration raised TAR (0.24 → 0.43, three-seed mean; not significant, p = 0.14) and, for seed 42, reduced activation on background structures (1.56 → 1.18). On YOLOv8-L and YOLOv9-C, the procedure defined model-specific zones and raised TAR (single runs), while detection gains came mainly from the schedule. The contribution is a validation-set procedure that bounds the λ search by TAR rather than by detection metrics and diagnoses background leakage, not an accuracy gain.