Research on a lightweight model for laser-cut diamond defect detection based on multi-module collaborative optimization.

As a high-end abrasive material, diamond is prone to defects such as cracks and ablation during laser cutting due to complex physical interactions. Achieving efficient and accurate detection of laser-cut diamond defects is crucial for enhancing product quality and production efficiency. Traditional deep learning object detection algorithms generally face challenges like large parameter magnitudes and heavy computational burdens, making them difficult to apply in potential scenarios of resource-restricted equipment such as handheld inspection devices. In response to this problem, this research develops FAS-YOLO, a lightweight defect detection model for laser-cut diamonds based on the YOLOv11n framework. Initially, the FDConv (Frequency Domain Convolution) module is integrated to reconstruct the C3k2 feature extraction component, enhancing the capture of core defect features.Second, the ADown (Adaptive Downsampling) module is employed to refine the downsampling layer, resolving parameter redundancy. The SEAM attention mechanism is integrated to enhance the detection head module. By learning the importance of different channels and fusing channel information, the model is guided to focus more precisely on defect region features while suppressing background interference such as metal reflections. Experimental results demonstrate that the FAS-YOLO model achieves 92% precision, 80.4% recall, and an mAP50 value of 82.6%. Compared to the baseline architecture YOLOv11n, it achieves competitive performance while reducing the number of parameters, GFLOPS, and model size by 37.4%, 40%, and 34.6%, respectively.