Architectural Design of RT-DETR-L for PCB Surface Defect Detection: A Systematic Comparison of Attention Mechanisms, Backbone Replacement, and Cross-Dataset Generalization

Abstract

Based on RT-DETR-L, this paper systematically compares five attention mechanisms (SE, CBAM, CA, ECA, and EMA) at the P3/P4/P5 outputs of the feature-pyramid neck under identical training conditions, and evaluates FasterNet backbone replacement and a P2 small-object detection head as complementary improvements. Experiments reveal a performance gap of up to 4.29 percentage points (CA: 93.04% to EMA: 97.33% in mAP₅₀), indicating that the choice of attention mechanism has a substantial impact on RT-DETR-type PCB detectors. EMA achieves the best mAP₅₀ (97.33%) and the highest mAP_50:95 (56.45%); ECA offers a competitive trade-off without increasing GFLOPs (96.69%); CA performs worst (93.04%), a 3.34 pp drop below the baseline, tentatively attributed to an architectural conflict with the AIFI encoder. FasterNet backbone replacement trades accuracy for efficiency (31% fewer parameters, 40% lower GFLOPs); and, when trained from scratch on the second dataset DeepPCB, the EMA variant again yields the largest gain (mAP₅₀89.33%, 4.83 pp over the baseline), showing that the improvement is not specific to a single dataset.