An Anti-vibration Hammer Detection Algorithm Based on Mobilenet V3 and YOLO V3

Abstract

Deep learning-based object detection algorithms have some limitations, such as complex network structure and high computing power requirements, which makes it difficult to meet real-time detection in transmission line inspection robots. In response to this problem, we propose an improved lightweight anti-vibration hammer detection algorithm based on MobileNet V3 and YOLO V3. As the collected anti-vibration hammer dataset is limited in size, Gaussian noise and salt & pepper noise are applied to expand the dataset, respectively, which are then fed into MobileNet V3 for feature extraction after labeled. Finally, the extracted features with MobileNet V3 are fed into the YOLO module for training. Experiments results show that the algorithm can detect anti-vibration hammer clearly and effectively, and the detection speed of 68ms per sheet can be achieved on low-power embedded devices.