Object Detection in Unmanned Aerial Vehicle (UAV) Images using YOLOv5 with Supervised Spatial Attention Module

There have been various object detection models developed recently that have enabled better results in object detection. Object detection has gradually increased in computer vision with further development of application areas. Unmanned aerial vehicle (UAV) images feature smaller and more fragmented objects, as opposed to landscape images where objects occupy more space. Furthermore, rotational and measuremental factors reduce object detection accuracy. In this paper, an improved object detection framework based on YOLOv5 is proposed in order to resolve these issues. As such, this study proposes SSAM-Darknet for the detection of UAV images based on objects. SSAM-Darknet and Bi-FPN are used to extract multiscale and multilevel features from the input images. Additionally, dilated convolution and the Ada-bound optimizer are employed to enhance the proposed model in detecting the objects from UAV images. This experiment evaluates the accuracy of object detection by using VisDrone-DET. AP (Average Precision) and AR (Average Recall) metrics are proposed as a quantitative way of evaluating detection performance. The proposed model achieves an average precision of 34.32 making an increase in detection accuracy by 10% compared to other detectors.