COVID-19 localization and recognition on chest radiographs based on Yolov5 and EfficientNet
Tong Zhang, Boxuan Zhang, Feiyun Zhao, Shiran Zhang
Abstract
Chest radiographs clearly present the characteristics of lung lesions in patients with new coronary pneumonia, thus they can be leveraged to build a new coronary pneumonia detection model to provide doctors with favorable auxiliary diagnosis results. This paper proposes a COVID-19 localization and identification approach based on yolov5 and EfficientNet. Due to inherent reasons such as computational complexity and network structure, the features of a single model are usually limited in representation, and EfficientNet provides yolov5 with competitive feature expression through BiFPN and other advantages, and the ensemble of EfficientNet and yolov5 recognition results will significantly improve the performance of a single model. In order to evaluate the robustness of the approach proposed in this paper, we trained our network on COVID-19 Detection datasets from Kaggle platform. Evaluations and comparisons demonstrate that the ensemble approach achieves better performance on various scenes.