Inspection AI Transforms from ‘Something to Buy’ to ‘Something to Create’

To get straight to the point:

A real-world dataset usable for industrial anomaly detection (inspection AI) has been made available for free. By combining this with few-shot learning, a pathway has opened up to build inspection AI that previously cost 3 million yen from an SIer for just 50,000 yen in-house.

The breakdown of the “50,000 yen” is simple. It includes the cost of a used PC with a GPU or cloud computing (a few thousand yen per month), a camera (which can be a few tens of thousands for industrial use), and the learning cost for engineers. The software and dataset are free. The bulk of the previous 3 million yen, which consisted of “SIer labor costs and custom data construction costs,” has been drastically reduced.

This is a significant event for small and medium-sized manufacturing companies.

What Has Been Released — The ‘ISP-AD’ Dataset Reflecting the Real Scene

The released dataset is the “Industrial Screen Printing Anomaly Detection Dataset (ISP-AD).” It is a large-scale dataset that includes actual defect images and synthetic defect images collected from the screen printing process.

There have been publicly available industrial datasets before, with the well-known MVTec AD being a staple in anomaly detection research. However, there was a problem: the lighting was uniform, the backgrounds were clean, and the defects were clearly visible. In other words, they were all “well-behaved” data.

The real-world situation is different.

There are lighting inconsistencies. Surfaces have subtle patterns. Defects have weak contrast and can be missed even by humans. ISP-AD reflects this “real-world reality”. By including weakly contrasted surface defects, subtle anomalies on structured patterns, and imperfect shooting conditions, the accuracy gap when applied in actual factories is minimized.

The issue of the discrepancy between accuracy in research papers and accuracy in the field has been the biggest pitfall in the introduction of inspection AI. ISP-AD directly addresses this.

Why ‘Few-Shot Learning’ Becomes a Weapon for Small and Medium Enterprises

Even with the dataset available for free, many small and medium enterprises do not have a large number of defective product images. While having fewer defective products is good, it poses a challenge for AI training.

This is where “few-shot learning” comes into play.

Traditional machine learning required thousands to tens of thousands of images. Few-shot learning can learn new anomaly patterns with literally just a few to several dozen images. By using a general dataset like ISP-AD to teach the base of “what constitutes an anomaly,” and then showing just a few images of their own products, detection models can start functioning.

Recent research has highlighted approaches utilizing Vision Foundation Models (VFMs). By incorporating an attention mechanism specialized for anomaly detection (Power-Law Self-Correlation Enhanced Attention = PL-SCEA) based on models pre-trained on large image datasets, it becomes possible to capture subtle changes in texture and structural distortions from a small number of samples.

There are even more advanced techniques. The framework called “Decoupling Product-Agnostic Anomaly Representations (DPA)” allows anomaly patterns learned from one product (such as scratches, dirt, or chips) to be applied to another product. In other words, “zero-shot” — even if there are zero defective images of their own products, there is potential to build an anomaly detection model.

This means there is no longer a need to collect hundreds of defective images for each product, which eliminates data collection costs.

Breaking Down the Cost Breakdown — What Consumed the 3 Million Yen?

When small and medium manufacturing companies outsource inspection AI to SIers, it typically costs between 2 million and 5 million yen. A typical breakdown is as follows:

It is important to note the “labor costs”. It will be necessary to have personnel in-house who understand the basics of AI. However, the level of “understanding” referred to here does not require the ability to read academic papers. A basic understanding of Python and the ability to run models following published tutorials is sufficient. Recently, environments have been established where models can be tested with near copy-paste operations using Hugging Face or Google Colab.

The majority of the previous 3 million yen was consumed by “labor costs for collecting and organizing data” and “development costs for building models from scratch.” These two costs are now approaching zero thanks to free datasets and pre-trained models. This is the essence of the structural change.

So, What Should We Do? — Three Actions Small and Medium Manufacturing Companies Should Take Now

1. Start Experimenting (Cost: 0 yen, Time: Half a Day)

Download ISP-AD and try running publicly available anomaly detection models (such as PatchCore or Anomalib) on Google Colab. Take a few images of your own products and input them. The accuracy can be rough. The first step is for someone in the company to experience “AI finding defective products.”

2. Quantify Your Inspection Costs (Cost: 0 yen, Time: 1 Day)

Identify the annual costs of defective products, the labor costs associated with visual inspections, and the costs of outsourced inspections. If this remains vague, it will be impossible to judge “what will change with AI.” If the annual cost of defective products exceeds 5 million yen, a 50,000 yen investment should be an easy decision.

3. Implement on a Small Scale (Cost: 30,000 – 50,000 yen, Time: 1-2 Weeks)

A used GPU-equipped PC (such as Jetson Nano for around 20,000 yen), a USB industrial camera (10,000 – 30,000 yen), and an open-source anomaly detection framework. This three-point set can build a prototype of inspection AI for one production line. Aim for perfection is not necessary. It’s sufficient to place it alongside as a “visual aid.”

This Is Not Just About Inspection

What is fundamentally happening is a structural change where “what could only be done by experts becomes accessible to anyone through the democratization of data and models.” Inspection AI is just one example of this.

The same structure extends to demand forecasting, inventory optimization, and predictive maintenance of equipment. The common factor is the multiplication of “the free release of high-quality datasets” and “the evolution of learning methods that operate with little data,” resulting in a dramatic change in implementation costs.

It is not inherently wrong to pay 3 million yen to an SIer. However, the time has come to question whether the contents of that 3 million yen “truly hold 3 million yen worth of value.” In an era where both data and models are available for free, what should we spend money on? The answer lies in “tuning tailored to our own site” and “systematizing operations.” This is where small and medium business owners should focus their thinking.

Local Small and Medium Manufacturing Companies Can Ride This Wave

Large companies are slow to move due to existing SIer contracts and internal approval processes. Small and medium enterprises can start as soon as the management decides to “do it.” The speed of decision-making becomes the greatest weapon in an era of advancing technological democratization.

In an age where inspection AI can be tried for 50,000 yen, there is no time to ponder over a 3 million yen estimate. Start experimenting. Quantify the numbers. Implement on a small scale. Just by doing that, the landscape of the inspection line will change.