To meet the growing demand for lightweight models and rapid defect detection in mini-light emitting diode (LED) chip manufacturing, we developed a highly efficient and lightweight multi-branch gradient backhaul (MBGB) block. Based on the MBGB block, a mini-LED surface defect detector was designed, which included an MBGB network (MBGB-net) for the backbone and an MBGB feature pyramid network (MBGB-FPN) for the feature fusion networks. MBGB-net was introduced to reduce resource utilisation and achieve efficient information flow while enhancing defect feature extraction from mini-LED wafers. MBGB-FPN optimises the parameter utilisation, thereby reducing the demand for computational resources while maintaining, or even improving, the detection accuracy. Furthermore, a partial convolution module is integrated into the detection head to reduce the computational overhead and improve the detection speed. The experimental results demonstrated that the method achieved optimal performance in terms of both accuracy and speed. On the mini-LED wafer defect dataset, it achieved an mAP50 of 87.2% with only 9.3M parameters and 21.6G FLOPs, reaching an impressive FPS of 345.4. Furthermore, on the NEU-DET dataset, an mAP50 of 77.5% was achieved using the same parameters and FLOPs.

中文翻译：

---

MBGB 探测器：用于 mini-LED 表面缺陷检测的多分支梯度回程轻量级模型


为了满足微型发光二极管 （LED） 芯片制造对轻量级模型和快速缺陷检测日益增长的需求，我们开发了一种高效、轻量级的多分支梯度回程 （MBGB） 模块。基于 MBGB 模块，设计了一种 mini-LED 表面缺陷探测器，其中包括用于主干网的 MBGB 网络 （MBGB-net） 和用于特征融合网络的 MBGB 特征金字塔网络 （MBGB-FPN）。MBGB-net 的引入是为了减少资源利用率并实现高效的信息流，同时增强 mini-LED 晶圆的缺陷特征提取。MBGB-FPN 优化了参数利用率，从而减少了对计算资源的需求，同时保持甚至提高了检测精度。此外，检测头中集成了部分卷积模块，以减少计算开销并提高检测速度。实验结果表明，该方法在准确性和速度方面都取得了最佳性能。在 mini-LED 晶圆缺陷数据集上，它实现了 87.2% 的 mAP50，仅 9.3M 参数和 21.6G FLOPs，达到了令人印象深刻的 345.4 FPS。此外，在 NEU-DET 数据集上，使用相同的参数和 FLOPs 实现了 77.5% 的 mAP50。