Detection of the Pin Defects of Power Transmission Lines Based on Improved TPH-MobileNetv3

Abstract

Pins are essential connecting components in power transmission lines. Their extensive use yet leads to frequent defects. Given the small size of a pin and many similar components, the detection of such defects is not ideal, which is a technological problem in the identification and diagnosis of power defects. In response to the large size, complex background, and on-site requirements, such as real-time detection, of power transmission lines, this paper proposes a method to detect pin defects based on TPH-MobileNetv3 (Transformer prediction Head Mobilenetv3). This paper modifies and adds a self-attention layer to MobilNetV3-Small to improve the feature extraction capability of small targets after downsampling. A feature fusion structure with layers of self-attention and a convolutional block attention module (CBAM) is added to the neck network, and a transformer prediction head are added to the head network so that different scale characteristics can be fused and focused from space and channels to strengthen the detection of small targets. Compared with the traditional MobileNetV3, the detection accuracy of the algorithm in this paper has been raised by 24%, as shown in the detection results of measured data. Moreover, compared with the mainstream algorithms with the same detection accuracy, this algorithm not only reduces the model size and significantly enhances detection efficiency but also satisfies the requirement of edge image processing of power inspection.