Object Detection in Urban Aerial Image Based on Advanced YOLO v3 Algorithm

Aiming at the characteristics of large changes in object scale and complex background in urban aerial image, we propose an advanced YOLOv3 detection algorithm to solve it. First, we analyze the aerial data through a clustering algorithm and calculate optimal size of the prior anchors. Then, a relatively lightweight and easily extensible backbone network-deep residual network is used for feature extraction. Meanwhile, in order to obtain a better receptive field and further reduce the information loss in the process of convolution, we add deformable convolution to the 8 times down-sampling feature map. And it improves the network's learning ability for geometrically deformed objects and attention mechanism before the 32 times down-sampling feature map to adaptively learn the weight relationship between different channels. Finally, IoU (Intersection over Union) is added to the regression loss function to improve the accuracy of evaluating the bounding box. The experimental results show that the inference speed of the optimized algorithm we proposed can reach 50 FPS, and the avera ge accuracy is 20% higher than the original network, which can complete the detection more effectively.