Thanks to the advanced use of Artificial Intelligence (AI) tools, Dimac is now capable of replicating even the most complex human visual inspections of metallic components.
This breakthrough extends quality control automation to sectors where traditional machine vision systems were previously unable to deliver reliable results.
The new generation of Dimac quality control machines for fasteners can now detect defects that were once impossible to identify using conventional systems.
The result is:
- more accurate and effective inspection,
- significant reduction of production waste,
- economic and energy savings, and
- a shorter time to market.
High-Speed Optical Inspection Powered by AI
Dimac sorting machines integrate high-speed optical control stations – up to 700 parts per minute – equipped with advanced cameras and lighting units, combined with AI-powered vision software. On the same machine, you can also perform checks with other traditional stations.
The AI software processes the grayscale images acquired from the cameras, interprets their content, and determines in real time whether each part is compliant or defective.
Dimac’s AI algorithms compare incoming images with a reference database, instantly identifying nonconforming parts.
Some application examples:
- The AI can differentiate an engraved marking from a surface crack, preventing false rejects.
- It can measure real dimensions even when the surface is dirty or contaminated.
- It can recognize appearance defects on materials with non-uniform color or finish.
AI does not merely “see” defects:
it understands them
Similarity-Based Algorithms and Continuous Learning
Unlike traditional rule-based vision systems, Dimac’s AI software does not rely on predefined parameters or thresholds.
Instead, the algorithm evaluates statistical similarity between analyzed images and reference categories, determining the probability that a new image belongs to either the “conforming” or “non-conforming” class.
Each “neural network” is initially trained to recognize a specific type of defect, but can be extended to inspect families of fasteners with similar characteristics (e.g., identical shape but different material).
Once trained, the model can be updated with new parameters or defect types while preserving the original training base.
The neural network is a file that is linked to the control plane and is automatically called every time the plane is loaded.
Dimac performs AI training as a Service
AI training requires:
- a dedicated high-performance computing station,
- the necessary software licenses, and
- specialized expertise in machine vision and machine learning.
The AI training process requires a suitable computing station, the necessary software licenses, and specialized skills, and can be performed remotely by Dimac technicians.
For further training, Dimac’s sorting machine onboard software includes a simple guided procedure to assist operators in capturing images of both good and defective samples.
In this way, Dimac’s engineers can carry out training or model updates remotely, in 3 days after receiving the images.
Starting from the images acquired on the machine, Dimac performs the AI model training, fine-tuning, and validation in-house, and then installs the trained neural network directly onto the customer’s sorting machine.
THE AI TRAINING PROCESS
1. Acquisition of Annotated Real Images
Real images of components are captured under various operating conditions (illumination, orientation, surface state), including both conforming and defective parts.
Defective areas are annotated manually or through semi-automatic labeling tools.
2. Data Augmentation
New image variations are created from real samples through geometric transformations (rotation, translation, zoom), as well as adjustments to brightness, contrast, and noise.
This enhances dataset diversity and improves model generalization.
3. Synthetic Image Generation
When real defect samples are rare or unavailable, Dimac employs photorealistic simulation, and procedural modeling to generate synthetic images of metallic components. Simulated defects include scratches, cracks, dents, or surface contaminations.
4. Domain Randomization and Controlled Variability
During synthetic image generation, controlled random variations in lighting, texture, camera angle, and noise are introduced. This technique, known as domain randomization, makes the model more robust to unpredictable real-world conditions and reduces the gap between synthetic and real data.
Beyond Fasteners: AI for Image Interpretation
The use of AI for visual interpretation has proven to be highly flexible and scalable.
The same algorithms developed by Dimac can be applied in any field requiring the human-like discernment of visual quality inspection — but with dramatically higher speed and a streamlined training process.
For example, these technologies are increasingly relevant in medical imaging and diagnostics, where AI can support the interpretation of complex visual data with exceptional precision.
Conclusion
The integration of Artificial Intelligence into Dimac’s sorting machines represents a major technological leap in automated quality control.
By combining high-speed vision, machine learning, and synthetic image simulation, Dimac pushes the boundaries of industrial visual inspection — delivering a new standard of reliability, precision, and efficiency.