Insulator detection method based on multi-angle region proposal network
Zhihao Chen, Yewei Xiao, Yan Zhou · 2020
Aiming at the direction of insulators are uncertain in the inspection images of transmission lines, and the problems of missed detection and inaccurate positioning due to complex backgrounds and occlusions, a method of insulators detection based on multi-angle region proposal network is proposed. Firstly, using a convolutional neural network, that characteristics of the different levels of features are obtained, and a feature fusion of shallow features and in-depth features is used to obtain richer features. Next, using the region proposal network based on Faster R-CNN, multi-angle factors are added and fine-tuned by non-maximum suppression so that it can generate candidate frames in any orientation to fit the target better. At the same time, the original pooling layer was changed to the RoIAlign layer. Finally, a 1 × 1 convolution layer is added before the fully connected layer of the network in order to reduce the parameters of its feature map and avoid overfitting. The inspection insulator image dataset is constructed and tested. The experimental results show that the final detection result obtained using this method can accurately detect the insulator, and its performance is better than the existing mainstream target detection algorithms and has high engineering application value.