| Bird strikes are one of the most critical factors threatening the production activities of airports,and intelligent and accurate acquisition of bird information is of great importance for the regular operation of airports.According to the situation of bird repelling in airports,we establish two main research tasks,bird object detection and fine-grained classification.The former aims to filter out the areas containing bird objects from the image,and the latter aims to distinguish high-confusing bird species accurately.For unique usage scenarios,we collect and organize two datasets dedicated to bird target detection and fine-grained recognition in airports.In order to prove the robust performance of the algorithms,some same publicly available standard benchmarks with similarity to both tasks are also adopted.For the airport bird object detection task,we propose the DSHYOLOV5.Based on the original YOLOV5,we add a decoupling header with Swin-Transformer.CBAM module and contextual module are also adopted.A copy&paste data enhancement is further designed to alleviate the small ratio of bird targets and the severe occlusion.The experimental results show that the DSH-YOLOV5 achieves an AP of 95.56%on the airport dataset and excellent performance results on the remaining public datasets.As for the fine-grained classification of airport birds,we propose the DA-VIT.Based on VIT,the relevance is fused with gradients to visually point out the high-response region of images and then use it as a benchmark to design three kinds of data augmentation:local area enlargement,flipping,and cutout.In order to widen the distance between confusing classes,a label smoothing mechanism is used to supervise the training process.According to the experiments,DA-VIT achieves 92.9%accuracy on the airport-based dataset with low computational cost,and four SOTAs are obtained on the remaining five fine-grained public datasets.Based on the above two algorithms,we design and implement a WEBbased application system,which realizes various functions such as user management,bird detection and fine-grained classification,and user feedback.The system has a user-friendly interface,easy to operate,and has a particular practical value. |