Insulator Semantic Segmentation in Aerial Images Based on Multiscale Feature Fusion

As one of the important components in the transmission line, the insulator is related to the safe and reliable operation of the entire transmission line. Aerial images are characterized by complex backgrounds, multiple pseudotargets, and low signal-to-noise ratios. Rapid and accurate localization of insulators in aerial images is a critical and challenging task in automatic inspection of transmission lines. Most insulator localization methods suffer from the loss of target edge detail information and large amount of model parameters. To solve these problems, this paper adopts an Encoder-Decoder architecture, called ED-Net, to realize end-to-end intelligent and accurate identification of insulators in aerial images. Firstly, Initial Module and CA-Bottleneck which are used to extract features from images to generate finer feature maps are proposed in the Encoder path. Meanwhile, global average pooling is used to preserve the maximum receptive field. Secondly, in the Decoder path, Refinement Boundary Module and Asymmetric Convolution Module are given to perform boundary optimization on the feature map, which are generated by the Encoder path. Finally, the Attention Feature Fusion Module is introduced into the Decoder path to combine high-level features with low-level features better and reduce the gap between features of different levels. The proposed model architecture keeps a suitable balance between the model parameters and insulator segmentation performance on insulator test datasets. Specifically, for a 512 × 512 input image, 95.12% mean intersection over union is achieved on the insulator test datasets with different environments and model parameters size being only 13.61 M. Compared with the current state-of-the-art semantic segmentation methods, the results show that the proposed method has higher efficient and accuracy.