Visual inspection automation by AI

"imaging", "judgment" and "operation"
Integrating three know-how
One stop solution! !

Do you have any of these problems?

I want to digitize and automate the work that depends on individual skills in the inspection process.

It takes a huge amount of time to design feature extraction (parameters) for inspection equipment...

The work of NG classification requires a lot of man-hours for inspectors, and we want to promote efficiency and automation.

Those who have concerns about AI utilization and are at a standstill

Appearance inspection is a very important process to meet high quality requirements from customers and to secure trust.
An increasing number of companies are working on automation with the aim of breaking away from individualized inspection methods.
Behind this is the improvement of imaging technology that replaces the human eye and the dramatic progress of AI technology that replaces the human brain.
In AI in particular, the era has changed from the era in which AI can distinguish things that can be distinguished by the human eye, to the era in which AI can distinguish things that cannot be distinguished by the human eye.
However, although these technological advances have increased the possibility of realizing automation, on the other hand, more specialized knowledge is required, and it has become difficult to proceed with the project with only the human resources of the company.

At our company, we have integrated the three know-hows of "imaging", "judgment", and "operation" that require expertise so that those who are going to work on automating appearance inspection can proceed with the project with peace of mind. We provide a one-stop service from the formulation of an overall plan to operational support.

In addition to automating the appearance inspection process, we analyze the causes of defects using acquired image data, and provide support for quality improvement solutions that identify defects and digitization of the entire manufacturing site.

Three things necessary for Visual inspection automation

AI visual inspection manual BOOK

This is a free material that explains in detail the points to consider: "imaging", "judgment", and "operation", which are the three essential elements for Visual inspection automation using AI!

Visual inspection automation,
Three of "imaging", "judgment" and "operation"
They are closely related.
just one of
Automation is not possible.

Visual inspection automation three elements image

Point of Imaging: Building an Environment That Can Detect Defects
Judgment point: Eliminate oversight
Operational points: data management, maintenance

Simply following the points above will not work.

Notes on the imaging phase
Just to take a picture of the target
tend to focus

[Failure example]

The total number of shots is enormous, and as a result, the tact time exceeds the specified limit.

[Solution example]

Succeeded in developing a model that does not miss anything by performing AI learning with an expanded imaging field of view

Notes on judgment phase
A model that detects NG only by the amount of learning
tend to rush to cover

[Failure example]

Accuracy improvement of specific NG detection hits a ceiling

[Solution example]

Succeeded in high-precision AI judgment by increasing lighting on the imaging side and performing specialized imaging for specific NGs that cannot be overlooked.

Points to note in the operation phase
We succeeded in device development and AI model development, but...

[Failure example]

Unable to secure human resources capable of managing and operating AI models, unable to upgrade...

[Solution example]

Succeeded in maintaining the operational level by formulating post-operational workflows in advance, planning countermeasures, and securing human resources.

Issues that may occur in each phase
There is a possibility that it can be solved by complementing each other.
In other words, after holding down each point,
Know-how in “imaging,” “judgment,” and “operation”
We need to consider it comprehensively.

1 Imaging

It is a very advanced technology that makes full use of optics, software technology and image processing technology.
As an example, there are the following methods.

  • A method to improve inspection performance by changing the lighting angle and lighting direction
  • A technique for reducing reflected light with a modified filter
  • Detection method by synthesizing images with different lighting directions
  • A method that separates colors for each lighting angle, such as coaxial lighting, ring lighting, and colored lighting, and detects them in batch imaging.

Imaging example 1 Engraved character inspection

Method: Detect by synthesizing images with different lighting directions
It can be seen that even engraved characters, which are difficult to discriminate in normal images, can be clearly discriminated.

Imaging example of engraved character inspection

Imaging example 2 Solder inspection

Method: Coaxial lighting, ring lighting, colored lighting, etc. Detect by batch imaging by classifying colors for each lighting angle
Imaging time can be shortened because of batch imaging.

Image to be inspected

Coaxial illumination image

Image to be inspected

ring lighting picture

Image to be inspected

Colored lighting image

Image to be inspected

Image of colored lighting

≪ Flow of introduction of standard imaging equipment ≫

*The order may change depending on the situation.

Inspection standard
confirmation of

  • We will ask you to present the criteria for defects that can be judged by a third party, such as defective items and judgment criteria.
  • You will be asked to present information on the parts that require imaging.

Mass production inspection process
Assumption of

  • We will confirm the inspection tact
  • Decide on imaging equipment and imaging method


  • Check if the defect shows up in the image
  • Assuming mass production, we will propose imaging conditions along with images.

imaging test

  • Check if defects can be detected from the image
  • Create a reproducible imaging environment, acquire images, and use the images for inspection.

Equipment introduction

Using the imaging mechanism created by image inspection, we will manufacture a visual inspection device that meets your needs.

≪Frequently asked questions about shooting≫

What is the difference between using an area sensor camera and a line sensor camera?

Line sensor cameras are expensive because the cameras and lenses are expensive, and the images must be taken while synchronizing with the actuators. Also, adjusting the camera is a time-consuming task. For this reason, we usually recommend area sensor cameras, which are inexpensive and easy to adjust.

2 Judgment

Here, we will introduce the points of introducing a judgment model by AI.

≪ Inspection method and characteristics ≫

We will introduce the general characteristics of AI image inspection, visual inspection, and conventional image inspection.
It is important to recognize and understand the characteristics of each in order to select the method that fits the problem to be solved.


AI image

imaging test

Ease of design


with roses


distinguishing features

consistency of judgment

Reliability of operation

processing speed


exact measurement

× ×

Inspection examples for which AI image inspection is easy to apply

Occurs with existing imaging tests
Reduction of gray zone (suppression of over-detection)

Dirt and scratches that cannot be dropped into rule-based inspections
Classification and sorting based on differences in characteristics, such as distinguishing between

Application example of AI image inspection Inspection of workpieces with uneven patterns

This is a foreign matter inspection for material-based workpieces that are difficult to distinguish with the human eye.
With current AI technology, it is possible to detect foreign objects with a high degree of difficulty like this.

Image to be inspected

Inspection object

Image showing AI reaction area

AI reaction point

≪ Flow of judgment model construction ≫

sample image

After concluding a non-disclosure agreement, we will keep sample images (OK/NG).

Data preprocessing

Consider the data structure by combining dozens of candidates for image cropping, reduction, binarization, etc.


Start learning by selecting tools, selecting networks, examining various parameters


We verify the accuracy of the AI model built by learning, consider it, and select the optimal model.
*Relearn if necessary

Accuracy report

We will report the accuracy of the developed AI model

≪Frequently asked questions about AI judgment≫

Should AI be procured or built in-house?

There are also companies that produce AI image inspections in-house from the viewpoint of internal maintenance of technology. However, the effort to keep up with fast-advancing AI technology is enormous. If you cannot secure engineers in-house, it is better to outsource AI technology.

What is the difference with existing inspection equipment?

AI is not omnipotent, and should be used according to its characteristics. We recommend that you consider "separation" with existing inspection equipment and "complementation by combination" according to the purpose and issues of introduction.

How do you evaluate the reliability of AI?

AI has been regarded as a "black Box", but recently, technology that visualizes the basis of judgment (e.g. heat map) has been established. It is desirable to use such technology to judge the reliability of AI.

How do you incorporate it into the line?

In order to utilize smart AI in factory inspections, it is important to properly incorporate it into the line. It is necessary to plan in advance who will incorporate it into the line and how, such as cooperation with other FA equipment and host systems.

3 Operation

In the operation phase, we will perform AI retraining to improve the accuracy of the judgment model.
Since AI relearning requires specialized knowledge, Macnica supports customers by providing AI operation services.

operational data

Full-scale collection of image data, which was scarce at the PoC stage

- AI Operation Service -

  • Provision of data storage system
  • Verification of AI model accuracy by experienced data scientists
  • Selection of data requiring retraining
  • AI model retraining
  • AI model implementation

AI model

AI learning using operational data

AI model

Implement the developed AI model

Accuracy verification

Link the results of regular human inspections and AI judgments


Relearn images that AI misses/overdetects

How to proceed with the Visual inspection automation project

When starting a project, Macnica first formulates an overall plan to address the client's unique challenges.
We also provide a variety of services at each phase to enable our clients to proceed with their projects with peace of mind.

How to proceed with the project


-Phase 1-
Overall plan

Grasping the current situation / Creating a plan / Examining cost effectiveness / Creating a work flow / Scrutinizing the estimate

Consulting Services

-Phase 2-
Imaging verification

Examination of imaging environment (camera, lighting, image size, tact time) / simple equipment estimate / rough estimate of operation equipment

・Imaging verification service
・Simple verification machine manufacturing service

-Phase 3-
Operation system
Requirement definition

Required specifications / business flow / functional requirements / system configuration / data storage system / production linkage system / quotation

Operation phase requirements definition service

-Phase 4-
Initial operation
System development

Imaging device development / transportation system / operation system introduction

・AI model development service
・ Imaging operation machine manufacturing service

-Phase 5-
AI development/
Fully automated

Operation image data collection / AI model development / AI model implementation

・AI operation service

Beyond automation

Manufacturing industry DX starting from Visual inspection automation

Various realizations can be obtained from the data (facts) obtained by visual inspection.
for example…

If there is data with many defects
→ Isn't it better to investigate and improve the cause of defects that occur in the previous process? the realization that

If there is data that the inspection quantity itself is decreasing
→ If the number of production changes, wouldn't it be better to establish a personnel shift? the realization that
→ Actually, if it turns out that there were many short stops, wouldn't it be better to do CBM? realization that

Awareness obtained from the data obtained in this way will be a stepping stone toward the realization of DX.

I know how to proceed! Seminar video

A seminar video showing how to proceed with Visual inspection automation is now available!
I want to hear more about the Visual inspection automation project! If you are interested, please listen to it.
With the cooperation of imaging partner Shindenshi Co., Ltd. and AI partner CEC Co., Ltd., we are delivering it in four parts.

  • [Introduction] Steps for visual inspection optimization
  • [Imaging] The most important step for automating visual inspection, the secret of building an imaging environment
  • [AI Utilization Edition] Appearance Inspection x The Forefront of AI ~Keep an Eye on AI's Actual Values and Possibilities~
  • [Automation, Beyond] The path of factory DX starting from the optimization of visual inspection

Related Documents

*Case leaflet*
Aisin AW Co., Ltd. Case study: Automating visual inspection of car navigation systems with AI

*Book article*
Smart factory and actual AI operation
*Application is required to download.

*Book article*
Issues and initiatives for AI utilization in manufacturing and manufacturing sites
*Application is required to download.