Detection of small target in aerial photography based on deep learning

In order to solve the problem of low recognition rate and poor positioning in aerial images, a target detection method based on deep learning is proposed. This algorithm uses VGG16 network as a fine tuning network and adds some deep network in it. Joint training is carried out by extracting the features of the shallow layers and the deep features of the target to overcome the contradiction between location and recognition in the process of detection. The singular value decomposition technology is used to compress the convolution features to reduce the computing and storage requirements of the model, and Multi scale training method is adopted to adapt to the change of aerial target scale. The experimental results show that 0.76 mAP can be implemented on the general data set PASCAL, and the detection speed is 16 fps. The 0.63 mAP can be achieved on the special aviation target data set UCAS-AOD, and the detection speed is 18 fps. It can satisfy the requirements for small target detection accuracy, and the detection speed can be close to the real-time detection effect.