Task-Driven Super Resolution: Object Detection in Low-resolution Images
This work addresses the challenge of enhancing object detection performance in low-resolution images for applications like surveillance or remote sensing, representing an incremental advance by optimizing SR and detection jointly rather than independently.
The paper tackles the problem of object detection in low-resolution images by proposing a task-driven super-resolution framework that integrates detection loss into SR training, resulting in consistent and significant accuracy improvements across various conditions and scaling factors.
We consider how image super resolution (SR) can contribute to an object detection task in low-resolution images. Intuitively, SR gives a positive impact on the object detection task. While several previous works demonstrated that this intuition is correct, SR and detector are optimized independently in these works. This paper proposes a novel framework to train a deep neural network where the SR sub-network explicitly incorporates a detection loss in its training objective, via a tradeoff with a traditional detection loss. This end-to-end training procedure allows us to train SR preprocessing for any differentiable detector. We demonstrate that our task-driven SR consistently and significantly improves accuracy of an object detector on low-resolution images for a variety of conditions and scaling factors.