Lu Wang, Zerui Wang, Yang Zhang, Jiaxuan Liang, and Li Liu
Insulator, object detection, multi-dimensional feature fusion (MFF), layer-interactive attention reinforced (LAR)
Insulators occupy a pivotal position in high-voltage transmission lines, and their effective detection emerges as a crucial component in preserving the stability and reliability of the power system. In the deep learning for insulator detection, the diverse angles and distances at which images are captured by unmanned aerial vehicles lead to challenges, such as insulator occlusion and small target issues. To address these issues, a multi-dimensional feature fusion and layer-interactive attention reinforced network (MLN) is proposed. Specifically, the MFF module can expand the receptive field by using parallel atrous convolutions with different sampling rates, and adap- tively construct convolutional kernels with different receptive fields to capture global feature information. It also achieves feature fusion in both channel and spatial dimensions, effectively solving the occlusion problem. The layer-interactive attention reinforced (LAR) module facilitates multi-scale feature learning by aggregating complementary information from adjacent features in both channel and spatial dimensions. It integrates deep and shallow feature maps to ensure that the model can simultaneously capture rich semantic information and spatial details, thereby improving the detection accuracy of tar- gets of different scales, especially small target objects. Experimental results show that the proposed model outperforms several current mainstream algorithms when trained on the proposed insulator dataset, achieving the best detection results with an accuracy of 0.940, a recall rate of 0.837, and [email protected] of 0.909. The dataset and source code are available on https://github.com/mi-luo/MLN.
Important Links:
Go Back
