TinyDefectNet: The little model that can help big manufacturers with better factory throughput
Factory production line throughput is only as fast as its slowest bottleneck. Sometimes the human element in quality inspection can slow the process down. Depending on the complexity of the product, number of inspection points, and accuracy of the human inspector, the inspection process can entail challenges for manufacturers.
As more factories integrate cameras and sensors into their production lines, AI can be added to automate relatively slow areas where humans are doing manual inspection, especially for tiny parts. New research from the University of Waterloo and DarwinAI illustrates how manufacturers can implement visual inspection with new levels of efficiency, spotting defects more quickly and accurately.
TinyDefectNet was specifically designed for industrial defect detection; it operates with an accuracy comparable to best-in-field models, but is 52x smaller and uses 11x fewer computational resources. It also benefits from our explainability-driven (XAI) performance validation to boost confidence that the decisions it makes are correct and trustworthy. This paper was presented at and won the Best AI Paper award at the Annual Conference on Vision and Intelligent Systems.
Traditionally, if you were going to build something yourself, you’d start with an off-the-shelf generic model like ResNet or MobileNet and manually train it for your specific job. This process is incredibly time-consuming, expensive, and the resulting model might never meet size and speed requirements for it to actually function in real world scenarios.
Machine-driven design exploration, such as what we used with TinyDefectNet, radically speeds up this process, allowing you to create a purpose-specific architecture in a fraction of the time that is smaller, faster, and better quality. In this study we created TinyDefectNet in conjunction with the ZenDNN accelerator library, which further improved its runtime latency.
A smaller and faster architecture allows for fewer storage requirements, less computational resources, and faster access. The closer a decision is made to real-time, the more efficient the process becomes. It is also more cost effective, by allowing one to run the system from smaller, less expensive, and potentially out-of-the-box hardware. And that’s just setting up in the factory. If one includes the manpower and resource savings from creating the model in weeks instead of years, the gains can be exponential.
This type of AI isn’t going to eliminate the need for human inspectors; but it will enable those inspectors to do their work faster and more accurately.
Research has always been an important aspect of life at DarwinA and we’re motivated to enable the widespread adoption of AI for smart factories as we enter the era of Industry 4.0. Efficient and powerful AI solutions, Like Tiny DefectNet, can truly be game-changers for the manufacturing industry. We will continue to push the boundaries of what is possible in this important area.
To learn more about DarwinAI visual quality inspection solutions, contact us.