Image classification tools built on deep learning

Convolutional neural networks (CNNs) are a class of deep learning neural networks, commonly used to analyze visual imagery. Machine vision software assesses examples of each image class and develops deep learning algorithms that can make inferences about the visual appearance of each class.

MIL X leverages deep learning to perform image classification. The latest addition is a coarse segmentation approach that maps image neighborhoods according to categories to identify and roughly locate the presence of specific features or defects. A second, global approach assigns images or image regions to pre-established classes. Image-oriented classification is particularly well-suited for analyzing images of highly textured, naturally varying, and acceptably deformed goods. A tree ensemble technique is used for feature-oriented classification, which categorizes objects of interest from features extracted from these images. 

Back to top

Broader toolset for 3D functionality

Classical 2D vision techniques are not always able to perform localization, recognition, inspection, or measurement tasks. That’s when 3D vision comes in, able to deliver accurate dimensional data for more complex imaging tasks.

With an expansive collection of tools for conducting 3D capture, display, processing, and analysis including metrology and registration, MIL X works with 3D data in the form of point clouds, depth maps, and/or elementary objects. Ensuring ease-of-use, all MIL X tools can use 3D data produced by profile and snapshot sensors, stereo and time-of-flight (ToF) cameras, STL, and PLY files. 

Back to top

Registration tools for HDR imaging

Sometimes it is necessary to fuse multiple images into a single option in order to bring out details not otherwise seen in the same lighting conditions.

MIL X offers a set of registration tools, including one for performing HDR imaging. Using multiple images of an identical scene taken at different camera exposure levels, this tool generates a single image that highlights a far greater dynamic range of luminosity than what would have been possible with a single illumination source. The greater range of luminance visible in the resulting image brings out details that otherwise would have been lost.

Back to top

Hands-on training and training-as-a-service options

To perform inferencing, CNNs must first be trained. Training involves building a dataset—including labeling images and augmenting the dataset with synthesized images—as well as monitoring and analyzing the training process.

MIL X offers users two distinct options for deep learning training of CNNs: Users can take a hands-on approach and train their own CNN, or, Matrox Imaging’s team of vision experts can perform the training on the users’ behalf. Whichever the route, MIL X provides the necessary infrastructure and interactive environment to build a training dataset and conduct different types of training, including transfer learning and fine-tuning.

Back to top

Prototyping environment to streamline development

Working with a programming library to assess the feasibility and best approach to developing a vision application can be both intimidating and time consuming.

MIL CoPilot, the companion interactive environment for MIL X, lets users experiment, prototype, and generate functional program code within its interactive environment. The latest update now offers training and inference support for users looking to train their own CNNs for deep learning analysis. Users can now label and augment required datasets, monitor the training process, and view results in clear, concise tables.

Back to top