ASBNet-DFD: Structural Implementation and TensorRT Deployment Characterization of a Lightweight PCB Defect Detection Network

Detecting tiny defects on printed circuit boards (PCBs) requires a balance among preserving high-resolution details, fusing semantics across scales, and maintaining deployment efficiency. This paper presents ASBNet-DFD, a lightweight defect detection network designed for this setting. Instead of repeatedly upsampling, downsampling, and concatenating features at their native resolutions, ASBNet-DFD concentrates shared semantic interaction from different feature levels in a fixed low-resolution space. The resulting shared context is then injected into three detection scales through scale-specific gates, reducing repeated cross-scale computation on high-resolution feature maps. Because the highest-resolution detection branch requires both local detail and semantic context, the network further introduces an asymmetric semantic bridge that provides a single directed semantic compensation from the adjacent medium-resolution backbone feature to this branch. This adds context for fine-grained detection while keeping the extra computation limited. With 640×640 input, ASBNet-DFD has 0.873 M parameters and requires 5.1 GFLOPs. Across seven independent training runs on DeepPCB, the validation mAP50–95 is 0.8109±0.0016, indicating low run-to-run variation under the fixed training protocol. Under a unified RTX 4070 and TensorRT FP16 deployment setup, the model achieves an mAP50–95 of 0.6919 on the independent test set with a network forward latency of 1.175 ms/image, lower than YOLO11n, YOLO12n, and YOLO26n under the same test protocol. Under the current quantization configuration, the INT8 mixed-precision implementation provides neither a speed benefit nor a storage benefit. These results show that, under the experimental conditions used here, concentrating most cross-scale semantic interaction in a low-resolution domain while retaining only the necessary directed semantic transfer to the highest-resolution branch can provide competitive detection performance and practical inference efficiency with a compact model.