Real Time Stereo YOLO Framework for Detection and Metric Sizing of Surface Defects in FDM 3D Printing (with YOLOv8, YOLOv11, and EFEN YOLOv8)

Abstract

Fused Deposition Modelling, or FDM 3D printing, is great for building all sorts of parts, but it’s not perfect. You often see surface problems—blobs, stringing, layer shifts—the kind of stuff that ruins part quality and wastes material. Right now, most people just check these defects by eye or turn to expensive scanners and CT machines. Either way, you don’t get real-time feedback, and you can’t measure defects in actual millimetres on the fly. This paper proposes the Stereo-YOLO Framework—a low-cost, camera-based system for inline FDM monitoring. Dual synchronized cameras enable stereo rectification and disparity mapping, while YOLOv8, YOLOv11, and EFEN-YOLOv8 detect defects in the left-view image. Detected bounding boxes convert to metric dimensions via depth-based triangulation. We set out to fix that. Using a custom dataset of 5,200 stereo image pairs from FDM prints, we put EFEN-YOLOv8 to the test. It nailed a 94.2% mAP@0.5 at 92 frames per second on an RTX 3060—outperforming YOLOv8 and YOLOv11 by 3 to 5 percentage points. Even better, when it came to measuring defects, it kept average length and width errors to just 0.33 mm, and area errors around 0.48 mm², all checked against caliper readings for defects up to 50 mm. This isn’t just about catching defects. It’s about measuring them, in real-world units, with regular hardware, right as the print happens.