Skip to main content
Using machine learning to optimize manual inspection

Using machine learning to optimize manual inspection

Artificial Intelligence is permeating an increasing amount of aspects of our life. It is also increasingly permeating all pores of the manufacturing industry, which enables to achieve more efficient production, and also a friendlier one towards customers and the manufacturing workers.

Quality control is considered one of the most critical manufacturing activities since it ensures the products conform to a set of requirements and specifications, building trust with the customer, boosting customer loyalty, and reinforcing the brand's reputation. One of the means to realize such quality control is through the visual inspection of manufactured products. Nevertheless, while performed manually, such inspection poses several challenges. First, operators' decision whether a product is defective or not is subjective, and certain variance exists based on the operators' training and experience and factors such as their tiredness. Furthermore, such an inspection approach has limited scalability (e.g., it requires training and scaling the number of inspectors proportional to the production scale). On the other side, automated visual inspection guarantees the same criteria for all inspected products while doing so efficiently and with minimal downtimes. Suppose such systems use machine learning to identify potential defects. In that case, a decision must be made up to what confidence level the models' outcomes are trusted to meet quality policies, deferring the rest of the products to manual inspection. Therefore, a hybrid approach can alleviate much of the human effort required for visual inspection while still relying on humans for more complex cases.

Manual visual inspection can be enhanced with artificial intelligence too. Generative Adversarial Networks can be used to generate synthetic images, to provide a stream of images with an equal part of good and defective products. This can force the operator to pay greater attention than when inspecting an imbalanced stream of images, where most images correspond to non-defective manufactured products. Furthermore, defect hints can be provided to the operators to help them find and label the defects more quickly, increasing the velocity and quality of the throughput.

For further insights on this topic, we invite you to read our paper "Towards a Comprehensive Visual Quality Inspection for Industry 4.0." which we presented at IFAC MIM 2022 conference. We will be glad to hear your thoughts!