Research and Implementation of Small Object Detection Algorithm for Power Embedded Devices

Currently in the power industry, there has been a demand for intelligent computing and real-time feedback at the edge using embedded devices. However, in some scenarios, the excessive computation still seriously affects the practicality of intelligent edge devices. Take transmission inspection drone aerial images as an example, the requirement to identify some small objects at large resolution makes the inference model can only be deployed on servers or desktops with strong computing power support. In order to realize the practicality of edge intelligence devices for small object detection, this paper adopts the strategy of hierarchical object detection. This paper first analyzes the problems faced by small object detection under constrained computation and proposes the strategy of using hierarchical object detection, then refines the design and validation of the sliding window selection for hierarchical object detection, and finally selects the base model to be used considering the accuracy, inference speed, and other factors. This enables fast detection of small objects such as missing pins on resource-limited embedded devices with good detection results.