An Adaptive Unsupervised Method for Powerline Aerial Image Quality Assessment
Qimeng Li, Zhang Zhi, Chenjun Sun, Xuechun Lv, Yangyang Lian, Jiao Xing · 2021
When helicopters or unmanned aerial vehicles are patrolling power transmission lines, due to the variability of camera angles and interference from sunlight, atmospheric environment, and other factors, the collected images are distorted, which adversely affects the subsequent image processing. To solve the above problems, a method for evaluating the image quality of transmission lines based on self-adaptively generated weights is proposed, and it is applied to image preselection. First, ResNet-50 (Residual Network-50) is used as a feature extraction network to extract deep semantic features; second, a weight generation network composed of convolutional networks is used to generate adaptive weight parameters; finally, the image and the weight parameters are sent to the quality score prediction network formed by the fully connected layer to generate a score map. The aerial image quality score predicted by the proposed method is highly consistent with human visual perception.