AI-Enabled Image Recognition System to Revolutionize the Manufacturing Line

In-house Practice/Oyama Factory

AI technology is dramatically changing production sites for manufacturing. AI enables manufacturers to automatically generate and revise image recognition system programs, which are used for assembling parts and various inspections, to respond quickly to changes in part specifications or manufacturing lines. By making intensive efforts on this technology for five years with Fujitsu Group companies, Fujitsu Laboratories has made visible progress in improving the productivity and QCD (quality, cost, delivery) of electronic parts manufacturing.

Automatically Generate Image-recognition Programs for Inspections

Electronic product manufacturing lines use image recognition systems in which image recognition systems help perform automated assembly (e.g., tasks for mounting and assembling components) and inspections throughout the manufacturing process.

However, these image recognition systems are increasingly becoming a bottleneck for the manufacturing industry to respond more flexibly and quickly to part specification changes and start-up of and changes to manufacturing lines.

To reduce this bottleneck, Fujitsu Laboratories has been automating image recognition systems. When changing a manufacturing line or part, the system must also be revised. The limited image data for validation bogs down developers with a lengthy trial and error process. However, AI technology can obtain the optimal solution from a small amount of image data.

So we decided to use specialized genetic programming. When tested on a parts assembly line, our image recognition system automatically generated code for inspection and achieved a nearly 100% recognition rate. Based on this success, the Fujitsu Group immediately started using AI technology at production sites.

Specialized Genetic Programming -- Producing Results at Production Sites

Though the image recognition system program's targets and purposes are set in advance, the amount of image data usable prior to operating manufacturing lines is limited. With a lot of data, deep learning could be used to generate an original algorithm, but here we need a different approach.

Specialized genetic programming can be configured with dedicated templates to shorten processing times and achieve high recognition rates. For example, templates can narrow it down to three processes: image enhancement, threshold process, and binary image handling.

The program evolves automatically by preparing training data from images of normal and defective parts to make pass/fail judgments, reliably delivering results.

Figure 1: Structure and optimization of the recognition program for template pattern matching

Reducing Development Time by 80%
while Maintaining Recognition Rates of 97%+

Fujitsu Laboratories uses specialized genetic programming jointly with Fujitsu's manufacturing division to achieve major results at production sites.

Our system can distinguish parts regardless of shape and location during the process of inspecting parts for misalignment. We reduced the time to develop programs for pass/fail inspection processes by 80%. Our parts assembly machines can now relearn as adjustments are made and still maintain recognition rates of 97%+. Positioning variations were cut in half, shortening work hours by 33%.

Going forward, our strategy is to offer this image recognition system as a cloud service. We are considering two paths for cloud services: to create a database and to provide services. We have done individual trials thus far, but our true desire is to transform all of manufacturing using the cloud. There are still many growth opportunities for applications of this system at production sites.