Coaxpress Frame Grabbers
Camera Link Frame Grabbers
Non-Standard AnalogFrame Grabbers
Standard PAL/NTSC/1080Pvideo capture cards
Image AnalysisSoftware Tools
Evaluation andprototyping applications
Image acquisition software
GigE Vision, USB3 Vision, CoaXPress
IMX Pregius, MIPI CSI‑2
Machine Vision Development Kit
Deep Learning segmentation library
Deep Learning works by training a neural network, teaching it how to classify a set of reference images. The performance of the process highly depends on how representative and extensive the set of reference images is. Deep Learning Bundle implements “data augmentation”, which creates additional reference images by modifying (for example by shifting, rotating, scaling) existing reference images within programmable limits. This allows Deep Learning Bundle to work with as few as one hundred training images per class.
Our “Fabric” sample dataset shows how the unsupervised mode of EasySegment can be used to detect and segment defects in Fabric with only a few good samples for training and no knowledge about what kind of defects are expected. Moreover, the unsupervised mode of EasySegment can be used to ease the annotation of the expected segmentation required for the supervised mode by reviewing and importing the results of the unsupervised mode as ground truth.
Neural Networks are computing systems inspired by the biological neural networks that constitute the human brain. Convolutional Neural Networks (CNN) are a class of deep, feed-forward artificial neural networks, most commonly applied to analyzing images. Deep Learning uses large CNNs to solve complex problems difficult or impossible to solve with so-called conventional computer vision algorithms. Deep Learning algorithms may be easier to use as they typically learn by example. They do not require the user to figure out how to classify or inspect parts. Instead, in an initial training phase, they learn just by being shown many images of the parts to be inspected. After successful training, they can be used to classify parts, or detect and segment defects.
EasySegment is the segmentation tool of Deep Learning Bundle. EasySegment performs defect detection and segmentation. It identifies parts that contain defects, and precisely pinpoints where they are in the image. The supervised mode of EasySegment works by learning a model of what is a defect and what is a “good” part in an image. This is done by training with images annotated with the expected segmentation. Then, the tool can be used to detect and segment the defects in new images. The supervised mode of EasySegment achieves better precision and can segment more complex defects than the unsupervised mode thanks to the knowledge of the expected segmentation.
Our “Coffee” sample dataset shows how the supervised mode of EasySegment can be used to efficiently detect and segment foreign materials on a production line, even when the foreign materials’ color and texture are very close to the product of interest.
EasySegment is the segmentation tool of Deep Learning Bundle. EasySegment performs defect detection and segmentation. It identifies parts that contain defects, and precisely pinpoints where they are in the image. The unsupervised mode of EasySegment works by learning a model of what is a “good” sample (i.e. a sample without any defect). This is done by training it only with images of “good” samples. Then, the tool can be used to classify new images as good or defective and segment the defects from these images. By training only with images of good samples, the unsupervised mode of EasySegment is able to perform inspection even when the type of defect is not known beforehand or when defective samples are not readily available.
Deep Learning is generally not suitable for applications requiring precise measurement or gauging. It is also not recommended when some types of errors (such as false negative) are completely unacceptable. The unsupervised mode of EasySegment is good for defect detection and segmentation tasks, especially when defectives samples are hard to come by. Deep Learning tools usually work very well with images of natural or manufactured objects that have complex surface patterns (e.g. wood, fabric, …) that make the detection of defects by conventional machine vision algorithm very hard. Besides, the "learn by example" paradigm of Deep Learning can also reduce the development time of a computer vision process.
Open eVision includes the free Deep Learning Studio application. This application assists the user during the creation of the dataset as well as the training and testing of the deep learning tool. For EasySegment, Deep Learning Studio integrates an annotation tool and can transform prediction into ground truth annotation. It also allows to graphically configure the tool to fit performance requirements. For example, after training, one can choose a tradeoff between a better defect detection rate or a better good detection rate.
Deep Learning generally requires significant amounts of processing power, especially during the learning phase. Deep Learning Bundle supports standard CPUs and automatically detects Nvidia CUDA-compatible GPUs in the PC. Using a single GPU typically accelerates the learning and the processing phases by a factor of 100.
To access this area, enter your email address and password below and click “SIGN IN”
Enter your email address and click “CREATE AN ACCOUNT”. You will receive a confirmation email shortly. Just click the link in that email to register.
It’s free and easy !