Improved YOLOv5 for Aerial Images Object Detection with the Introduction of Attention Mechanism

Due to the development of drone technology and researches on deep learning in computer vision, object detection of aerial images from UAVs has a wide range of applications in practical production. Aerial images have problems such as large scale variations, small and dense objects and background disturbance. Traditional object detection methods are difficult to apply in practical detection scenarios, while deep learning-based methods each has its pros and cons. In this paper, the YOLOv5 model, which is a good fit for UAVs, is chosen to deal with the challenges of aerial image object detection, and the CBAM attention module is introduced to improve the network of YOLOv5 to obtain better feature extraction capability. This paper uses dataset analysis to obtain insight into the characteristics of objects in different datasets, and uses DOTA dataset for training, combining image pre-processing and data augmentation techniques to improve the training process and enhance the robustness of the model. In addition, this paper uses Grad-CAM visualisation method in an attempt to analyse and interpret the detection results. Compared to the original YOLOv5, the improved model has an accuracy rate of 75.8%, gaining a 2.2% increase. This shows the model performs well and the attention mechanism can successfully assist in resolving the challenging issues with object detection in aerial photography scenarios.