SBIR/STTR Award attributes
The National Image Interpretability Rating Scale (NIIRS) characterizes the quality of remote imagery and provides a means to directly relate the quality of an image to interpretation tasks. Since the advent of deep learning (DL), insufficient research has been performed to gauge the suitability of state-of-the-art machine learning (ML) models for NIIRS tasks. The proposed effort aims to construct a dataset spanning NIIRS 4 to 6 imagery, assess transferability of state-of-the-art object detection models operating on this data, and demonstrate practical image enhancement techniques to improve detection performance on NIIRS 4 imagery. Additionally, a hyperparameter optimization methodology will be developed to improve task performance. The salient aspects of the proposed solution are: (1) a repository for organization of collected datasets; (2) a transfer learning module for streamlining the evaluation and transfer of object detector layers; (3) a super-resolution module geared toward improving image interpretability; and (4) a meta-optimization module for automatic hyperparameter configuration. In Phase I, modules will be developed independently then integrated into a holistic framework. Feasibility will be demonstrated via case studies of US Air Force interest. The Phase II effort will focus on dataset expansion, algorithm optimization, extensive technology validation, and technology insertion into Air Force workflow.
