MIL includes Classification tools for automatically categorizing image content or previously extracted features using machine learning.
Image-oriented classification makes use of deep learning—specifically convolutional neural network (CNN)—technology in two distinct approaches. The global approach assigns images or image regions to pre-established classes. Results for each image or image region consist of the most likely class and a score for each class. The coarse segmentation4 approach generates maps indicating the pre-established class and score for all image neighborhoods. Image-oriented classification is particularly well-suited for analyzing images of highly textured, naturally varying, and acceptably deformed goods.
Users can opt to train a CNN on their own4 or commission Matrox Imaging to do so using previously collected images; these images must be both adequate in number and representative of the expected application conditions. Different types of training are supported, such as transfer learning and fine-tuning, all starting from one of the supplied pre-defined CNN architectures. MIL provides the necessary infrastructure and interactive environment to build the required training dataset—including the labeling of images and augmenting the dataset with synthesized images—as well as monitoring and analyzing the training process. Training is accomplished using a NVIDIA GPU or x64-based CPU while inference is performed on a CPU, avoiding the need for specialized GPU hardware.
Feature-oriented classification4 uses a tree ensemble technique to categorize objects of interest from their features, expressed in numerical form, obtained from prior analysis using tools like Blob Analysis. The categorization is made by majority voting of the individual-feature decision trees. As with image-oriented classification, users can train a tree ensemble on their own using the facilities provided in MIL or employ Matrox Imaging for the task.