Online Recognition of Defects and Foreign Objects Based on Wavelet Scattering Convolutional Networks

During the inspection of transmission lines by unmanned aerial vehicle (UAV), the captured images are typically transmitted to the cloud for processing, and currently, online detection cannot be performed locally on the UAV. This paper proposes a lightweight model based on an improved wavelet scattering convolutional network that enables local online identification of defects and foreign objects on edge devices. This lightweight model comprises an improved wavelet scattering network and a small convolutional neural network (CNN). The improved wavelet scattering network is more flexible than the classical network and can effectively extract the key features of the picture; The small CNN is divided into two parts: a four-layer convolutional network section for further extraction of small target features, and a three-layer fully connected for feature recognition and classification. Experimental results demonstrate that this lightweight model exhibits strong generalization and real-time performance, with accuracy rates for defect and foreign object identification exceeding 90%. The recognition speed is three times faster than that of YOLOv5s and nearly four times faster than YOLOv8s. The recognition speed on various edge devices exceeds 30 frames per second (FPS), meeting the 30FPS of UAV-captured videos, and reaching up to 153.8 FPS on Orange Pi 5B, enabling local online detection frame by frame.