Real-Time Foreign Object Recognition Based on Improved Wavelet Scattering Deep Network and Edge Computing

TL;DR

This paper presents a lightweight, real-time foreign object recognition model using an improved wavelet scattering deep network suitable for edge devices, achieving high accuracy and fast inference for UAV safety applications.

Contribution

The authors develop a novel, efficient wavelet scattering-based deep network that outperforms YOLO models in foreign object recognition on edge devices.

Findings

Recognition accuracy exceeds 90% on edge devices.

Inference time is less than 7ms for 720p images.

Model outperforms YOLOv5s and YOLOv8s in accuracy.

Abstract

The increasing penetration rate of new energy in the power system has put forward higher requirements for the operation and maintenance of substations and transmission lines. Using the Unmanned Aerial Vehicles (UAV) to identify foreign object in real time can quickly and effectively eliminate potential safety hazards. However, due to the limited computation power, the captured image cannot be real-time processed on edge devices in UAV locally. To overcome this problem, a lightweight model based on an improved wavelet scatter deep network is proposed. This model contains improved wavelet scattering network for extracting the scatter coefficients and modulus coefficients of image single channel, replacing the role of convolutional layer and pooling layer in convolutional neural network. The following 3 fully connected layers, also constituted a simplified Multilayer Perceptron (MLP), are used to classify the extracted features. Experiments prove that the model constructed with biorthogonal wavelets basis is able to recognize and classify the foreign object in edge devices such as Raspberry Pi and Jetson Nano, with accuracy higher than 90% and inference time less than 7ms for 720P (1280*720) images. Further experiments demonstrate that the recognition accuracy of our model is 1.1% higher than YOLOv5s and 0.3% higher than YOLOv8s.