Performance Analysis of Roboflow 3.0 and YOLO-NAS in PCB Defect Detection Compared to YOLOv9 and RT-DETR Using Augmented Image Dataset

The printed circuit board (PCB) market is increasing exponentially over the years, currently having a compound annual growth rate of 4.28% in Asia. Ensuring the quality of PCB’s is essential, as they are present in almost all electronic products. This research analyzed the performance of the YOLONAS and Roboflow 3.0 models, provided by Roboflow, in detecting defects in PCB copper. A public dataset of $\mathbf{9, 0 1 3}$ augmented images from Peking University was used. Subsequently, the results obtained were compared with the latest options in the neural network market, being the YOLOv9 and RT-DETR models. From a comparative analysis of results, the Roboflow 3.0 model stood out with a mAP of 99.20%, accuracy of 98.10%, recall of $\mathbf{9 9. 3 0 \%}$ and a mAP50:95 of $\mathbf{7 8. 5 0 \%}$. It outperformed YOLOv9 only in accuracy by 0.8% and mAP50:95 by 0.2%.