High-Accuracy Low-Latency Method for UAV Inspection Image Inference and Offloading

DNNs can significantly improve the accuracy of identifying key components and detecting defects in power line inspection images. However, the inference process of DNNs requires a considerable amount of computing resources and energy-constrained UAVs for power line inspection cannot achieve the accuracy and real-time requirements. This paper introduces CFOD, a precision technique with minimal latency for image analysis and data transfer in UAV-based power line inspection. CFOD includes a feature restoration module (FRM) to recover low-level features from high-level features, reducing communication overhead. Concurrently, CFOD’s adaptive collaboration scheduler (ACS) can dynamically choose the appropriate partition point for inference tasks according to system data and transfer a segment of the DNN model to the cloud server. Furthermore, our approach uses JPEG algorithm to compress the data transmission size of intermediate feature maps to further reduce latency. Experimental results show that compared to edge-only and traditional partition methods, our approach can save 46% and 33% of the inference latency, respectively.