Abstract: Infrared small dim target detection is a key technology that restricts the monitoring of unmanned aerial vehicle(UAV) at a long distance. To solve the challenges, such as less available target features and slow processing speed, we studied the network structure and designed a new target feature respectively. In terms of network structure, a detection architecture based on the fusion of empirical feature and deep feature was proposed, which took empirical feature as the starting point to promote lightweight neural network to attentively learn more deep features. In terms of new target featur, based on the analysis of the real target images, in addition to the energy significance, the local morphological closure feature of small targets were proposed. Besides, the local maximum difference convolution and corresponding residual convolution were designed, which were aimed to local morphological closure features. The experiment results based on the real data show that the detection rate of the proposed method can reach more than 95%, and the processing speed can reach 60 frame/s (FPS) . So, the performance and speed are better than the comparison methods, and the balance between detection performance and processing speed can be obtained. This work provides a new idea to put the infrared small dim target detection based on neural network into the real application.