What is automated defect detection?

Automated defect detection uses cameras and AI to inspect products and surfaces for flaws, scratches, cracks, dents, missing parts, and contamination, in real time, with no manual checking. [Ultralytics YOLO](https://www.ultralytics.com/yolo) models detect, localize, and classify each defect in a single pass as parts move down the line.

How accurate is YOLO for defect detection?

It depends on the data, but recent peer-reviewed studies report 84-99% accuracy with YOLO models tuned to a specific task, for example [99.4% on fabric](https://doi.org/10.3390/s25072270), [98.8% mAP on PCBs](https://doi.org/10.1038/s41598-025-20387-x), and [98.2% on semiconductor wafers](https://doi.org/10.1088/1361-6501/ade27d). The [Ultralytics Platform](https://platform.ultralytics.com/) supports active learning, so accuracy keeps climbing as new defect types appear.

Which industries use YOLO for defect detection?

Published 2024-2026 studies apply [Ultralytics YOLO](https://www.ultralytics.com/yolo) to steel and metal surfaces, printed circuit boards, semiconductor wafers, welds, textiles, solar panels, aluminum castings, ceramics, batteries, and pharmaceutical packaging, anywhere visual quality control matters.

Can YOLO catch defects that traditional machine vision misses?

Yes. Rule-based machine vision struggles with variable, subtle, or previously-unseen defects. [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) learns from your own examples to recognize scratches, hairline cracks, white spots, and missing components across changing lighting and part variation, reaching 98%+ mAP on several published benchmarks.

How fast is YOLO defect detection?

Fast enough for the line. [YOLO26](https://docs.ultralytics.com/models/yolo26/) processes a frame in about 1.7 ms on a modern GPU; published inspection systems report [186 frames per second on fabric](https://doi.org/10.3390/s25072270) and [2.3 ms per weld](https://doi.org/10.1007/s11554-025-01730-x).

Can defect detection run offline on the production line?

Yes. Most inspection deployments run on-premises because production data is sensitive and factory networks are unreliable. [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) exports to 19 formats, including TensorRT for NVIDIA Jetson, OpenVINO for Intel CPUs, and Hailo accelerators, so models run line-side with no internet connection required.

How much training data do I need?

Less than you might expect. Published defect-detection models train on a few thousand labeled images, for example the bridge-crack study used a 4,029-image dataset. The [Ultralytics Platform](https://platform.ultralytics.com/) handles annotation and training, and pretrained YOLO weights mean you fine-tune rather than start from scratch.

How do I get started with YOLO defect detection?

Collect images of your good and defective parts, annotate them, and train, all in one place on the [Ultralytics Platform](https://platform.ultralytics.com/). For production deployment, [enterprise licensing](https://www.ultralytics.com/license) covers commercial use and keeps your code and data private.

Automated defect detection

Catch defects the moment they appear. Across recent peer-reviewed studies, YOLO-based inspection reaches 99%+ accuracy and cuts manual inspection costs by up to 94.5%, on steel, PCBs, fabric, solar panels, and welds. Train it on your own defects and deploy it on the line.

世界有数の組織から信頼されています

How Ultralytics YOLO tackles defect detection

One model, every defect type

Ultralytics YOLO detects, localizes, and classifies defects in a single pass, and researchers have pushed it past 98% mAP on benchmarks from PCBs to steel to wafers. That means fewer escaped defects, less scrap, and faster time-to-market, on the hardware you already run.

Subtle defects, caught: 98.8% mAP on tiny PCB defects, 98.2% on wafer surfaces (peer-reviewed).
Line-speed inference: YOLO26 runs a frame in ~1.7 ms; weld studies report 2.3 ms per part.
Trainable on your defects: fine-tune on your own images in hours; published models hit 99%+ accuracy.
Deploy anywhere: export to 19 formats — TensorRT, ONNX, OpenVINO, CoreML — for edge or cloud.