Edge Computing Driven Low-Light Image Dynamic Enhancement for Object Detection

With fast increase in volume of mobile multimedia data, how to apply powerful deep learning methods to process data with real-time response becomes a major issue. Meanwhile, edge computing structure helps improve response time and user experience by bringing flexible computation and storage capabilities. Considering both technologies for successful AI-based applications, we propose an edge-computing driven and end-to-end framework to perform tasks of image enhancement and object detection under low-light conditions. The framework consists of a cloud-based enhancement and an edge-based detection stage. In the first stage, we establish connections between edge devices and cloud servers to input re-scaled illumination parts of low-light images, where enhancement subnetworks are dynamically and parallel coupled to compute enhanced illumination parts based on low-light context. During the edge-based detection stage, edge devices could accurately and rapidly detect objects based on cloud-computed informative feature map. Experimental results show the proposed method significantly improves detection performance in low-light conditions with low latency running on edge devices.