Abstract: In order to improve the detection and recognition effect of the two-stage target detection algorithm Faster R-CNN on small target of drones, an improved Faster R-CNN target detection algorithm is proposed. In the improved algorithm, firstly, the feature extraction network of the original Faster R-CNN algorithm is improved, and ResNet-18 with fewer convolutional layers is used as the backbone network to reduce the number of parameters of the algorithm. Secondly, according to the characteristics of the drone target, the feature fusion method in the feature pyramid networks of Faster R-CNN is improved to enhance the contrast between target feature and background feature. Thirdly, the bilinear interpolation method is used to solve the problem of the deviation of the prediction frame caused by the pooling of regions of interest. Furthermore, the verification experiments are carried out on the constructed low-altitude drone data set. The results show that the improved Faster R-CNN target detection algorithm has a detection speed of 35.5 frames per second (FPS), which is about double the speed of the original Faster R-CNN algorithm (15.8 FPS), and the improved algorithm’s mean avergage precison (mAP) is increased by 0.7%, which effectively improves the detection and recognition performance of the algorithm for drone small targets.

Key words: image processing, target detection, deep learning, drones