OBJECT DETECTION IN LOW-SPATIAL-RESOLUTION AERIAL IMAGERY USING CONVOLUTIONAL NEURAL NETWORKS
Chapman, Brannon W.
Scrofani, James W.
MetadataShow full item record
Supervised machine learning by convolutional neural networks has proven effective for regional object detection in digital imagery. However, when applied to low-spatial-resolution aerial imagery, such networks are generally less effective because of the low object-to-image size ratio, the unconstrained orientation of objects, and a shortage of labeled data. The purpose of this research is to assess whether a deep learning technique can be optimized for regional detection and classification of ships, aircraft, or similar platforms in aerial imagery. During tests, we sought and observed improvements in detection precision resulting from adaptations to the region proposal technique of the Faster Regional Convolutional Neural Network (R-CNN) model while using a shallower, fourteen-layer network. Specifically, we found that both k-means clustering and a segmented least-squares fitting technique reveal object orientation patterns in training data that can be used as the basis for the dimensions of Faster R-CNN region proposals. Detection precision was most notably improved for objects not tightly bound by a rectangular region due to their orientation in the image plane.
Approved for public release. distribution is unlimited
Showing items related by title, author, creator and subject.
Kanaev, A.V.; Daniel, B.J.; Neumann, J.G.; Kim, A.M.; Lee, K.R. (2011-10);Data fusion from disparate sensors significantly improves automated man-made target detection performance compared to that of just an individual sensor. In particular, it can solve hyperspectral imagery (HSI) detection ...
Downs, Justin (Monterey, California: Naval Postgraduate School, 2017-03);Given the problem of detecting objects in video, existing neural-network solutions rely on a post-processing step to combine information across frames and strengthen conclusions. This technique has been successful for ...
Finite element modeling and long-wave infrared imaging for detection and identification of buried objects Tilley, Heather P. (Monterey, California: Naval Postgraduate School, 2017-12);Detection of buried improvised explosive devices (IED) represents a complex threat to U.S. forces. This thesis explores the potential use of infrared images combined with finite element models to detect buried objects in ...