4 Tools for Deep Learning Factory Automation Inspections
Deep learning software automate a vast assortment of production functions that are impractical for human workers and rule-based algorithms.
Consider the inspector working on an automotive production line: doors, fenders, seats, windows, and hundreds more components can get scratched, dented, ripped or chipped along the way. Humans can catch some of those defects. Machine vision systems with high-tech cameras and complex algorithms can flag a few more well-defined, predetermined flaws.
The trouble is that all the variables in a production setting can produce imperfections that are impossible to anticipate. That is where deep learning software comes to the rescue: It uses digital cameras and image recognition algorithms to learn to identify a broad spectrum of problems like rust, discoloration, and damage.
When developed properly, deep learning applications help manufacturers reduce errors and improve product quality. In an inspection application, machine vision and deep learning work together like this:
- Developers build a base of training images of undamaged products to establish the “correct” product appearance.
- Developers add images of damaged products to identify the most common anomalies and flaws.
- Machine vision cameras take pictures of items on production lines. The machine learning application compares these new images against the training images to flag possible defects.
- Because the application is optimized to seek success and avoid failure, it essentially teaches itself to become more accurate over time.
The software to build deep learning applications for manufacturing must have four core capabilities:
1. Feature Location and Assembly Verification
Finding flaws isn’t the only role for machine vision and deep learning software. It also can use training images and learning algorithms to locate specific components, which can help with tasks like instructing a robotic arm to align components properly. This is essential for high-precision products like semiconductors, smartphones and pharmaceuticals.
These applications also can scan the number of products in a location and tell a robot to keep adding more of the same products until a shelf or carton is full. They also can count all the components in a package to ensure that nothing has been left out.
The best location and verification tools work in a wide variety of lighting and surfaces that confound rules-based vision systems and quality-control personnel.
2. Defect Detection and Segmentation
Identifying defects is perhaps the most sought-after capability for machine learning software in production environments. While machine vision systems can be programmed to flag one kind of flaw, identifying multiple flaws in this way is far too time-consuming.
Defect-detection tools start with a base of “good” images and pictures of common flaws like rust, dents, scratches and misalignments. Top-quality detection tools also have an option for identifying any flaws that depart from the “good” image. These images of rare production outcomes can help the tool teach itself to improve its accuracy.
Segmentation identifies one section within an image, telling the software to scan that area for flaws. This helps simplify deep learning applications by filtering out areas that aren’t relevant to the segment scan.
3. Object and Scene Classification
Classifying objects and scenes helps deep learning applications divide flaws into classes, which helps optimize the application’s ability to self-improve without human intervention. In general, images are labeled according to certain characteristics and then classified according to specific parameters. That way, for instance, scratched products could be automatically rerouted to the paint line while dented products could be sent to the metalworking shop.
Classification also sorts products and components based on common characteristics like color, texture, materials, packaging and defect type. The best classification tools establish tolerances for natural deviations in shades, shapes or dimensions, and vary these tolerances according to the needs of each class.
Read more: How deep learning classification tool works
4. Text and Character Reading
Reading words, numbers, or text consistently on a surface like an engine block or copper tube can be all but impossible for people and standard machine-vision algorithms. Lighting can vary widely on a production line, crating shadows in some places and glare in others — and shifting throughout the day depending on changes in ambient light on the factory floor.
Deep learning applications connect fonts and typefaces with the lettering on parts in production. This makes it much easier to read text through plastic covers and on uneven surfaces like clothing or gardening tools. Advanced character reading tools transcend the factory floor, finding a role in sophisticated distribution, logistics and commerce systems.
Other Features to Look for in Deep Learning Software
In addition to the four capabilities outlined above, a robust deep learning software package should be:
- Easy to learn, with an intuitive GUI that does not require advanced technical knowledge.
- Optimized for visual-inspection production environments, with smaller image sets that require less training.
- Designed for Windows PCs with GPUs (graphics processing units).
Cognex Deep Learning has these and many more powerful features designed precisely for factory and production environments, unlike other open source deep learning frameworks. It combines a comprehensive machine vision tool library with advanced deep learning tools inside a common development and deployment framework.