Data Labeling and Curation at Scale (DLCS) for Machine Learning Algorithms

Homeland Security operations create a large amount of data from millimeter-wave and Computed Tomography X-ray detection systems.The Transportation Security Administration’s Electronic Baggage Screening Program provides an excellent example of the immense volume of screening conducted and X-ray image data generated, since the Aviation and Transportation Security Act (Pub. L. 107-71, 2001) requires that 100% of checked baggage be screened at airports.Machine Learning (ML) for image detection can become an integral part of the screening process in order to further automate the inspection, expedite the process, significantly reduce the human involvement and allow baggage inspection to be extended in other modes of public transportation.A very large number of labeled images must be available for training ML algorithms.The more labeled images are available the better the training will be.Creating the training images requires a prohibiting large manual effort.The scope of the proposed project is the development of a Data Labeling and Curation at Scale (DLCS) process for creating curated training data with minimal manual effort.An instance discrimination approach will be used for training the feature extractor using the available unlabeled images.A You Only Look Once (YOLO) iterative algorithm with a fast manual review will be developed.It will be using the feature extractor trained by the instance discrimination and it will train the segmentation and labeling capabilities of YOLO.The DLCS will be capable of accessing data stored using HDF or DICOS formats.