Electrical Fittings Inspection Based on Improved Unet with Generative Adversarial Network and Attention Mechanism
Zhihui Xie, Min Fu, Xuefeng Liu · 2023
Inspecting electric power fittings in transmission lines is a complex task prone to errors due to the challenges posed by the variable shooting angles and the complex background in the images. Additionally, there exists the issue of mutual occlusion that occurs between fittings and imbalanced data sample categories. These challenges can result in false detections and leaks in the fittings. Therefore, a fittings detection method based on an improved U-shaped network (Unet) with a fast generative adversarial network and attention mechanism (FA-Unet) is proposed in this paper to address the aforementioned challenges. Firstly, virtual fittings samples are produced by the generative adversarial network to address the issue of unequal sample size in the dataset. Next, a mixed attention mechanism that combines self-attention and convolution (ACmix) is adopted to improve the model's ability to extract fittings features in complex backgrounds. Finally, an enhanced mix loss function is developed to solve the problem of unbalanced positive and negative samples in the fitting dataset. The experimental results demonstrate that the FA-Unet has achieved a commendable detection performance. The approach considers the network's performance and improves the detection effect compared to the original Unet.