A Detection Algorithm Based on Improved Cascade R-CNN for UAV Aerial Images

Due to the large change of target scale in UAV aerial images, the down-sampling method adopted by the traditional Cascade R-CNN backbone network will lead to the loss of small target semantic information, which leads to the reduction of target detection accuracy. This paper proposes a small target detection model for UAV aerial images based on improved Cascade R-CNN, the HRNet-W40 model is added to Cascade R-CNN, it reduces the loss of key information that may be caused by up-sampling and down-sampling, and effectively improves the detection effect of small-scale targets. Finally, the experimental results on the dataset collected by AISKYEYE show that the improved Cascade R-CNN obtains higher accuracy compared with other advanced UAV small target detection algorithms with the same number of parameters and computation, compared with adding other backbone networks to Cascade R-CNN, the mAP is improved by at least 11.7%, and compared with other advanced UAV small target detection algorithms, the mAP value is improved by at least 3.6%..