Introduction
In the wave of modern smart manufacturing and Industry 4.0, automated product quality control is the lifeline of factories. From tiny scratches on the surface of electronic components (PCBs) to machined holes on automotive metal parts, achieving high-precision, high-efficiency Automated Defect Detection has always been the core and trickiest subject in the field of industrial machine vision.
This article outlines the evolution of defect detection technology from “traditional rule-based image processing” to “data-driven deep learning,” and explores the engineering challenges faced in real-world deployments.
1. Traditional Machine Vision: The Art of Light and Rules
Before the full explosion of deep learning, defect detection on industrial sites heavily relied on traditional image processing algorithms. You can think of this workflow as doing extremely rigorous feature engineering, which usually includes the following steps:
- Image Preprocessing: Color-to-grayscale conversion, histogram equalization, and denoising (Gaussian filtering, median filtering).
- Edge Extraction and Feature Separation: Using Canny or Sobel operators, or separating foreground from background via morphological operations (dilation, erosion, opening, closing).
- Rule Calculation and Threshold Matching: Setting thresholds to determine whether a defect exists based on geometric features like area, circularity, and aspect ratio calculated from Blobs.
Pros and Cons
- ✅ Pros: Entirely white-box logic, highly interpretable; extremely low computing power requirements, capable of achieving very high real-time performance (usually in milliseconds) on ordinary CPU industrial PCs.
- ❌ Cons: Heavy dependence on lighting hardware (bright field, dark field, low-angle ring light). If the background texture is complex (like brushed metal), or if defect morphology undergoes unpredictable changes, algorithms relying on manually set thresholds will completely fail, causing severe “missed detections” or “over-kills.”
2. The Deep Learning Era: A Data-Driven Paradigm Shift
With the evolution of CNN (Convolutional Neural Network) architectures, industrial defect detection gradually shifted to being dominated by “data engineering.” Depending on varying production requirements, mainstream deep learning solutions have evolved into three major branches:
2.1 Object Detection
Classic algorithms like the YOLO series (YOLOv8, YOLOv10) or Faster R-CNN. The network can directly draw bounding boxes around defects in the input image and provide a confidence score.
- Applicable scenarios: Assembly lines where defect features are obvious, categories are diverse, and fast inference speeds are required.
2.2 Semantic Segmentation
If you need to know not only “where” but also the “exact shape and area of the defect” (e.g., to determine if the actual length of a scratch exceeds standards), pixel-level prediction models like U-Net or Mask R-CNN are necessary. In segmentation networks, to counter the severe class imbalance problem of “huge background, tiny defect,” we usually introduce improved loss functions:
$$ L_{total} = L_{ce} + \lambda \left( 1 - \frac{2 | X \cap Y |}{|X| + |Y|} \right) $$
(The formula above shows the combination of Cross-Entropy loss and Dice Loss. With the KaTeX rendering engine, such theoretical derivations can be perfectly presented on the blog.)
2.3 Unsupervised Anomaly Detection
This is currently the hottest frontier in industrial AI. On actual production lines, our biggest pain point is “numerous good products, extremely few defect samples” (or even an inability to collect defect images at all during new product introduction). Unsupervised algorithms represented by PatchCore and PaDiM have changed the approach: Train the model using only normal good product images, allowing the model to learn the feature distribution space of good products. During inference, any image region that deviates from this normal distribution is judged as an anomaly (defect).
3. Deployment Challenges: From Lab to Factory Floor
Although algorithms dominate open datasets (like MVTec AD), in actual deployments, algorithm engineers must overcome the engineering obstacles of the “last mile”:
- Data Silos and Incremental Learning: Factories are extremely protective of data privacy. Models need to support rapid fine-tuning and incremental updates on local edge devices with few samples.
- Limited Compute and Edge Inference: Assembly line cycle times are typically less than 50ms. How to achieve this speed on power-constrained edge devices (like NVIDIA Jetson or basic x86 IPCs without GPUs)? This requires extreme computational graph optimization through ONNX conversion, INT8 Model Quantization, and the use of inference engines like TensorRT / OpenVINO.
Academic & Design Insights
- Design Philosophy: In industrial vision, we advocate for “hybrid intelligence.” Instead of blindly pursuing complex models, we should choose the most robust solution based on operating conditions.
- Technical Breakthrough: By using traditional morphological operators as pre-filters for deep learning, we significantly reduced the computational overhead on the edge.
- Inspiration: When facing industrial requirements, algorithm engineers must understand that “Cycle Time” and “Stability” carry more weight than mere mAP scores.
Conclusion
Technology is never absolutely good or bad, only whether it fits the business scenario. In stamping parts detection with fixed morphology and single backgrounds, traditional morphology remains the most efficient and low-cost king. However, on complex fabric textures or reflective solder joint surfaces, deep learning is the only path to break through the ceiling. The excellent industrial vision systems of the future will inevitably be a deep integration of rule-based processing and AI neural networks.