No two eggs are alike, but they are all edible. This principle applies equally to industrial production. This means that the variety and potential defects of manufactured parts are virtually limitless. Are there additional challenges in quality assurance and identification tasks? As with eggs, deviations are often tolerated. That is, not all deviations are defects. This is where deep learning technology comes into play, enabling previously unimaginable image processing applications.
Comprehensive automated defect inspection is essential in all manufacturing departments. Increase customer satisfaction and trust while ensuring strict quality standards, minimizing scrap rates and ultimately reducing costs. On the other hand, manual inspection processes are significantly more time-consuming, prone to human error, and do not take advantage of the benefits of digitalization.
Automating inspection workflows is therefore a strategic investment, delivering maximum accuracy and throughput while freeing up human operators for more complex tasks. Machine vision has emerged as an important enabler for automated quality assurance, providing faster, more reliable, and more robust results than the human eye in most industrial scenarios. The entire workflow is accelerated, achieving inspection cycle times of less than 20 milliseconds per component.
Most machine vision systems that have been in use for many years operate on a rule-based approach. This means that certain algorithms must be programmed to follow predefined rules in all possible application scenarios. These methods allow various tasks to be solved quickly and efficiently, allowing companies to significantly improve their operational efficiency. This has been the reality in the world of machine vision up until now.
Advances brought about by artificial intelligence (AI) mark the beginning of a new era in machine vision. AI, particularly its subset known as deep learning, breaks through the limitations inherent in rule-based methodologies.
Open the door to next generation machine vision applications
The most innovative benefit of deep learning is that it reduces programming effort. Instead of hard-coding all defect types, the system is automatically trained using a large set of representative image data. This self-learning capability enables previously unfeasible machine vision applications.
Using the food industry as an example, natural products can be highly volatile and even change over time, yet still be viable for consumption. Deep learning-based solutions can reliably distinguish between OK and NOK products despite extreme differences that rule-based algorithms cannot achieve. Techniques such as anomaly detection ensure accurate classification without false negatives.
Another use case is the inspection of weld seams, which is similarly highly variable. Deep learning-based classification enables automated and reliable quality assessment when the underlying neural network is trained with acceptable weld image data.
Software can streamline data processing, labeling, and training for such applications. Known defect classes (such as flaws, air entrainment, and underwelds) can be defined and trained to accurately classify them.
Beyond inspection: advanced automation scenarios
Deep learning also optimizes complex bin picking and pick-and-place operations. For rigid CAD-based objects, deep 3D matching enables 3D bin picking using only 2D image data. For easily deformable or translucent objects, such as plastic bags filled with assembly parts, methods such as object detection and grasp point detection allow robots to reliably grasp objects, even when stacked or oriented in random directions. Robust performance is achieved through extensive image-based training to accommodate virtually infinite shape and position changes.
Another area powered by deep learning is optical character recognition (OCR). Traditional OCR suffers from reflections, surface irregularities, and uneven lighting. Deep OCR maintains high recognition rates under these conditions, accurately locating characters regardless of orientation, font, or polarity. It also groups letters into words, eliminates misinterpretation of visually similar symbols, and thanks to pre-trained deep neural networks, it significantly improves accuracy even on hard-to-read text.
Proven in real use cases
These technologies are integral to our standard machine vision software and deliver tangible benefits across a variety of industries. Another notable real-world example comes from the automotive sector. A global manufacturer of luxury car batteries leveraged the manufacturer’s Vision software to automate the entire cell assembly process, including visual inspection and packaging. As a result, defective products were prevented from leaving the factory, false negative classifications were reduced by 57.3%, and both quality and efficiency were improved.
Bridging deep learning and rule-based machine vision to enable new applications
Machine vision, and deep learning in particular, is proving essential to modern quality assurance across all manufacturing sectors through its wide range of methodologies and use cases. Its accuracy and fast performance reduce costs, optimize resource utilization, and maintain customer satisfaction through consistent product quality.
That said, both rule-based algorithms and deep learning approaches have their place in industrial automation. The most innovative machine vision applications leverage the best of both paradigms. Combining these technologies allows manufacturers to achieve maximum robustness and throughput in automated inspection and processing tasks.
Advanced machine vision platforms allow users to seamlessly access and integrate rule-based and deep learning-based methods within a single workflow. This hybrid approach allows you to automate new applications with maximum speed and robustness. And in the end, as with all eggs, we ensure that all the parts fit in the right place inside the packaging box.
