Aiming at the problem of occlusion and confusion targets recognition for UAV (unmanned air vehicle) under small sample conditions, a target recognition model integrating self-attention mechanism and few-shot learning was proposed. On the basis of using the idea of meta learning to obtain the ability of few-shot learning, the self-attention mechanism to learn the context dependence between the internal parts of the target was introduced into the model, so as to enhance the target representation ability and solve the problem of insufficient effective features in the case of occlusion and confusion. In order to verify the effect of the model, the occlusion and confusion target datasets were constructed by further processing the reference datasets and UAV aerial photography data, and different occlusion degrees and background confusion rates were set. Through the verification on different datasets and compared with the deep learning model, the proposed model is proved to possess higher learning efficiency and recognition accuracy.