Deep Inspection

An AI image comparison inspection system to support a variety of industriesDeep Inspection

When it comes to decision-making, existing machine and image evaluation systems fail where humans succeed. Deep Inspection was developed with that challenge in mind.
We created an extremely high performance system with high-speed uniformity, wherein the more experience is gained, the more accurate the evaluation becomes.


With Deep Inspection, the AI recognizes the part of the image that matches a particular pattern and determines whether that pattern meets the criteria. “When it comes to decision-making, existing image processing systems fail where humans succeed.”
By using deep learning to deal with such challenges, the AI is able to flexibly sort patterns like the human mind would.


Decisions made using Rist’s original “confidence score”

The confidence score is calculated by making a comparison with past results. Those with high scores are assigned automatic inspection, while those with low scores are left to be inspected by the human eye. By making it so that only those with low confidence scores are examined by people, manual labor is effectively reduced, dramatically increasing the inspection accuracy of the whole factory.

Improvement through continuous operation

Humans improve with experience, and so does deep learning. As learning data grows, the AI’s judgments increase in accuracy. An AI that has built up experience through continuous operation becomes a system that can be relied upon like a seasoned craftsman.

Can be installed into existing lines

Our product is not a package, it is a system that makes use of the customer’s current production line. With the capacity to design for a variety of different work sites, we offer products with long-term operation and upgrades in mind.

Mirror surface inspection process

case.01Mirror surface inspection process


Rist developed mirror surface inspection systems using Deep Learning for the product inspection process at Murakami Corporation, holder of the top market share for back mirrors in Japan. This move improved inspection accuracy (from 60% to 97%), and also reduced the strain on inspectors carrying out inspections visually.

Problems with conventional systems

When numerical values such as the color of the image to be inspected are digitized, the standard is ambiguous because people decide the threshold value.

Around 60% accuracy

In the end, the product must be examined by the human eye.

Rist’s Approach

Uses multi-class Convolutional Neural Networking (CNN)

Multi-class Convolutional Neural Networking (CNN)

Development of Rist’s original “Deep Inspection Confidence Score”

Deep Inspection Confidence Score
97% accuracy achieved
  • After installation on all lines, the number of inspection workers was reduced by 70%
  • With the goal of overseas expansion in mind, data obtained through automated inspections is being used to optimize upstream processes
99% accuracy achieved
Package inspection process based on area extraction

case.02Package inspection process based on area extraction

We consulted with major package label manufacturers on product inspection accuracy with their current inspection devices. With Deep Learning utilizing regional extraction networks, accuracy improvements 100 times greater than conventional inspection systems were achieved.

Problems with conventional systems

Because of shadows and wrinkles, mistakes are made when attempting to determine the area to be measured.

About one error per 100 packages measured

Rist’s Approach

Utilizing regional extraction networks used in autonomous driving and medicine

Fusion Net

Extract and measure area with packages using AI

Extract and measure area with packages using AI
100 times higher accuracy achieved

Installation Process

  1. 01Examination and Analysis

    The installation method is planned according to the end goal, whether it is new construction, or implementation into the customer’s products or existing production line.
    Feasibility will be examined based on a rough consultation that could occur via email, videoconference, or face to face. Feel free to get in touch with us and start the process.

  2. 02System Design

    After collecting data specific to the customer’s worksite, we devise an optimal learning algorithm, and then go through prototype development before entering main development and delivering the product.When necessary, we collaborate with companies in various fields to design and develop hardware than can handle special inspections.

  3. 03Result Measurement and Verification

    With the constructed line in action, we measure and verify result data, fine-tuning the AI program to optimize its performance.

  4. 04Beginning System Operation

    After installation, we continue to support customers through maintenance inspections and additional AI learning.
    We also offer consultations for more advanced inspections.