Real-time Computational Enhancement of Video Streams

In this SBIR Phase I project, we introduce an innovative video enhancement solution designed to significantly improve video quality in airborne systems. Our approach is built upon an integrated deep learning pipeline that tackles deblurring, super-resolution, and artifact removal simultaneously. The first component in our deep learning pipeline is a Convolutional Neural Network (CNN)-based classifier that can accurately identifying various types of image degradation. This classification enhances the efficiency of our multi-task video enhancement architecture, which is designed with parallel branches to address different enhancement challenges. An attention fusion module will be used to combine features from each branch. Additionally, our system incorporates optical flow estimation to maintain temporal consistency across video frames and capture the dynamics of moving scenes. Another major focus of our development is ensuring our video enhancement pipeline operates within Size, Weight, and Power (SWaP) constraints, which is important for practical deployment. To achieve this, we will implement advanced optimization techniques, including pruning and quantization, to streamline our model without compromising performance. Our proposed technology will not only advance the state-of-the-art in video enhancement technology but also solve challenges in achieving high-quality video in airborne systems.