AI tools like ChatGPT were trained using open source materials. What if your AI application needs domain-specific data that is hidden behind a wall of secrecy? In this article, Miron Shtiglitz, VP of Product and Delivery at visual inspection company QualiSense, explains the importance of domain-specific data in developing AI applications for manufacturing.
One of the fundamental drivers behind the effectiveness of deep learning models is the availability of large amounts of labeled data. Although obtaining such data is becoming increasingly easier for various applications, significant hurdles exist in developing AI-driven systems for production environments, such as quality inspection systems.
The concept of pre-trained deep learning models is gaining traction as a means to facilitate rapid development across multiple applications, and we are even hearing about deep learning models that require only minor tuning to perform optimally in some applications.
These models have a fundamental understanding of important features and are able to distinguish and make sense of complex information.
The availability of huge image datasets online has played a key role in fueling the growth of these models. Datasets such as ImageNet, Coco, NuScense, Google Open Dataset, etc. cover a wide variety of scenarios, from animals to nature images, object detection, and more. These datasets can be an invaluable starting point, especially for applications in that domain. However, challenges arise when there is not an abundance of publicly available data in the domain. This is especially true in the industrial domain. For example, image datasets of industrial processes are often limited as they are not provided by manufacturers, making it more difficult to develop models in this domain.
Industrial Quality Inspection
Consider an application that is outside the scope of a traditional dataset, a situation commonly encountered in quality inspection: using a pre-trained model can provide a slight advantage in some situations, but this advantage diminishes as the application moves away from the domain of the dataset.
For example, if your application deals with grayscale, hyperspectral, or LWIR imagery, a pre-trained model designed for color imagery may unintentionally hinder your work by extracting irrelevant features that don't match your specific domain. These challenges highlight the need for domain-specific data in the development of deep learning applications.
In a relatively conservative manufacturing environment, it's understandable that companies are hesitant to share their proprietary data, which may hold secrets about the complex processes and inspection solutions that set them apart from their competitors.
QualiSense has risen to this challenge by forming a strategic partnership with Johnson Electric, a global leader in the automotive industry. This partnership gave us access to a wide variety of production lines, giving us a treasure trove of imagery in the industrial inspection domain. These images encompass a wide variety of processes, materials, and applications. This unique repository of data gives QualiSense a competitive advantage. We used this data to build AI models that have a general understanding of the production domain, but can adapt to the complexities of each specific production environment.
The rise of deep learning has highlighted the importance of data in various forms. In quality inspection and other manufacturing applications, the lack of accessible unique data is a major challenge. While pre-trained models have a role to play, the real key to success lies in domain-specific data that reflects the complexity of the production environment.
For more opinions and insights on the technical challenges facing AI-driven quality inspection, check out the QualiSense blog.: qualisense.ai/blog
