CREATIVE PERCEPTION: Grey Matters’ Semi-supervised detection remote sensing prototype algorithm suite

The National Geospatial Intelligence Agency (NGA) requires a state-of-the-art model to quickly find and define new target sets within ever increasing datasets. Grey Matters’ CREATIVE PERCEPTION will provide the NGA with an advanced detection algorithm suite that can be easily upgraded to new target sets with a minimum amount of labeled training data. CREATIVE PERCEPTION, based on a complex self-supervised learning technique to extract useful high-level representations from high-dimensional data, will dramatically reduce the training data required to reliably identify low-shot or rare objects in remote sensing imagery to give the NGA the competitive advantage they require.  Grey Matters will design and demonstrate methods that integrate unlabeled data with small volumes of labeled data to augment our detection performance in panchromatic EO imagery. Grey Matters’ novel solution to solving the low-shot detection problem is to design a deep neural network (DNN) architecture that understands the semantic meanings behind class labels. These semantic meanings and relationships of class labels are learned through Natural Language Processing (NLP) techniques to convert the class labels into an embedding representation. A DNN is then trained to classify images to this embedding space as opposed to a simple one-category label output. Once the network has learned this new image to embedding space mapping, the network can provide an embedding for a new target class with just a single example. Objects can then be identified in unlabeled data by that new target embedding.