Airport Near-Altitude Flying Birds Detection Based on Information Compensation Multiscale Feature Fusion

Airport near-altitude flying birds’ detection is a critical task to avoid collisions of bird-to-aircraft. However, there still are two challenges as follows: 1) the inconspicuous features of birds make the detection task challenging, due to the small sizes and 2) deep networks are unfriendly to edge devices, while lightweight networks may exacerbate information loss due to depth separable convolution (DSC). In this article, we propose a lightweight network, named as ICFF-YOLOv5, to address the aforementioned problems and detect near-altitude flying birds effectively. To avoid the feature loss, we design a module of feature fusion, named as information compensation feature fusion (ICFF) to reduce the semantic difference between feature maps of different scales. Furthermore, we develop a double combination convolution (DCC) to reduce the number of parameters and achieve an advanced performance. In addition, we construct a dataset to evaluate the performance of our method. The experimental results demonstrate that the propose method can accurately detect the flying birds with varying shapes and scenarios, and achieve the values of evaluation metric mAP0.5 with 82.3 and F1 score with 82.4.