Convolutional Neural Network · CNN
Convolutional Neural Network
What is a convolutional neural network?
- Captures an image’s local features, layer by layer
- Builds up from lines to shapes to objects to recognize images
- The workhorse of computer vision
What CNNs changed is 'who creates the features.' Before, engineers hand-designed formulas that measured edge direction or changes in brightness, and when the product changed, they rewrote those formulas from scratch. A CNN turned the filter values themselves, the equivalent of those formulas, into something to be trained. The nature of the work therefore shifted to swapping in new data, not new code, when the shop floor changes, and the development bottleneck became getting labeled images, not the algorithm. In practice, this is the difference between '6 months of algorithm development' and '6 weeks of data collection.'
The structure usually splits into convolutional layers, pooling layers, and fully connected layers that make the final call. Because a single 3×3 filter sweeps the whole image, the number of values to learn drops by a factor of several hundred compared with a fully connected network of the same size, and the same filter responds whether the object is on the left or the right of the photo. In practice, training from scratch is rare. Transfer learning is the default: take a model pretrained on 1.2 million ImageNet images and retrain only its final layers. With a few thousand in-house images, that is usually enough. Vision transformers (ViT) have recently pulled ahead on large datasets, but CNNs remain the stable choice for in-house datasets of a few thousand images.
The most common mistake is putting a model on the line based on validation accuracy alone. In a process with a 1% defect rate, labeling everything as good still yields 99% accuracy, so the number you should actually watch is recall, which measures how many defects you catch. Also, a CNN is fairly robust to shifts in position but is not automatically robust to rotation, scaling, or changes in lighting color temperature, and detection rates often plummet after switching to night lighting. If consecutive frames of the same product end up split across the training and validation sets, data leakage produces scores far higher than real performance, and the true performance only shows after the model goes onto the line. There is also a common misconception that you cannot see the basis for the model’s decisions, but tools like Grad-CAM show a heat map of which part of the product the model looked at when it called it defective.
Ara LimExpert in global leadership, AI, and business communication
Seungtae ParkAX practitioner who teaches AI business development and marketing automation by building it hands-onIts foundations were laid by Yann LeCun’s research in 1989, and AlexNet’s 2012 competition win ushered in the deep learning era.
Consider a hypothetical case. Company A, an automotive injection-molded parts maker with 240 employees and annual revenue of KRW 78 (≈ US$1) billion, had six inspectors check appearance by eye. Screening 30,000 parts a day by sight let scratches slip through, leading to 4 customer claims a year, each costing KRW 20 million (≈ US$15,000) in full-lot sorting alone. The quality team proposed adopting CNN-based visual inspection as a three-month project.
- Director Lee on the quality team dug through three years of inspection records, but found only 1,200 labeled images of real defects. There were 280,000 photos of good parts.
- Manager Kim in production engineering said the imaging jig had to come before any training. "The lighting angle is different every day. What is there to train on?" he asked, and spent two weeks fixing the illumination and the distance.
- An outside vendor fine-tuned a pretrained ResNet-50 on 1,200 defect images, augmented with rotation and brightness shifts to 9,600, and reached 97% validation accuracy in three weeks.
- At the review meeting, the CEO asked, "How many defects did we miss?" and 18 undetected cases came to light. The team decided to lower the decision threshold based on recall, accepting more false rejects.
- Six months later, customer claims for escaped defects fell from 4 a year to 0, and 4 of the 6 inspectors were reassigned to re-inspection and packing. In return, people double-check the 3.2% of false rejects.