RCVNet: A bird damage identification network for power towers based on fusion of RF images and visual images

The technology for identifying birds around power towers using cameras alone is still susceptible to environmental interference. This paper proposes a new bird damage recognition network, RCVNet, which addresses this issue by fusing radio-frequency (RF) images and visual images. The network employs a feature layer fusion approach that accurately identifies bird damages in the monitoring area. Initially, RCVNet takes a group of RF and visual images as input. Then, through a series of convolutional neural networks (CNNs), birds are identified and located. To overcome challenges in recognizing small targets, several improved modules such as crosssupervised fusion network (CSF-net), posture deformable convolution (PDF), small-target attention fusion mechanism (SAFM), and Tiny-YOLOHead are introduced throughout RCVNet, improving surface information utilization and small feature retention rates. Finally, a bird damage discrimination strategy is developed based on the recognition outcomes of birds. As there is currently no public dataset available for RCVNet training, a new bird dataset called CRB2022, which includes RF and visual images, was gathered. Through large-scale experiments utilizing these methods, RCVNet effectively identifies birds, achieving a mean average precision of 79.34% and a mean average recall of 83.29%. Additionally, the discrimination rate of the utilized strategy can reach up to 98%.