Abstract: Anti drone system is an effective way to identify and attack the "black flying" drone. Image recognition drone is one of the keys of anti drone system. Aiming at the problems that the samples collected from drones are small samples, the features are not enough and the recognition accuracy is not high enough, an image recognition method of anti drone system based on transfer learning, dense convolutional network and coordinate attention mechanism was proposed. Firstly, a variety of drone images in different backgrounds were collected by using self-made device, and data samples were set up; Secondly, the network TL-CA4-DenseNet-121 based on transfer learning, coordinate attention mechanism and dense convolutional network, the network TL-SE4-DenseNet-121 based on channel attention mechanism were designed to identify small samples. The designed network was used to identify small samples and compare. The network recognition experiment of coordinate attention module and channel attention module based on different positions and different numbers were carried out respectively; Finally, the network with the best recognition effect was compared with the classical convolutional neural network models. The experimental results show that the proposed TL-CA4-DenseNet-121 network has better recognition effect than other networks, and the average accuracy of recognition is 97.93%, F1-Score is 0.9826 and training time is 6832 s. It shows the superiority and feasibility of this network in identifying small sample drones.