Ian Seremet

Conformal 3D printing techniques enable the deposition of inks on non-planar surfaces, but the presence of air pockets within inks can result in under-extrusion errors. This study lays the groundwork for automating dataset creation and labeling for training a YOLO object detection model to identify under-extrusion defects in images of conformal 3D prints. Data was produced by toggling an extruder for specified durations, and ten images were captured of each layer from multiple orientations. These raw images underwent pre-processing to extract only the printed material. Specifically, training data was selected using an automated cropping tool, which divided an image into a grid of windows. Windows that exceeded an average red pixel intensity threshold were saved as part of the training dataset. The images were then manually annotated using ‘Computer Vision Annotation Tool’ (CVAT) to highlight under-extrusion defects with bounding boxes. Labels and images were exported and reformatted to ensure compatibility with the YOLO v11 model. An initial training was conducted with a dataset of 219 images yielding a mean average precision (mAP) of 0.01. Expanding this dataset to 1,000-5,000 images is likely necessary for moderate model performance, requiring 20-25 additional prints to be photographed and processed. Rotation and noise augmentation may assist the model in handling nuanced image cases. Furthermore, CVAT's automated labelling feature could be harnessed to streamline the annotation process for the larger dataset. With these enhancements, a high-accuracy model capable of detecting under-extrusion defects becomes attainable.