A new single-stage deep convolution detection network is proposed to solve the complex background problem of night light remote sensing. Firstly, a classification network is designed by extracting high-dimensional features and then selecting features, and the influence of different channel number networks of noise reduction is studied. A prior box matching of gray-scale energy is proposed, inputting a low-noise and high-quality matching box into SSD detection network, and the idea of integral diagram is used to simplify the calculation. By adding sequential connection and dense connection to improve the global semantic module, the cross layered information interaction of the network is introduced, and its attention map comprehensively considers the high and low receptive fields to effectively distinguish small targets and background noise. Experimental results of the night light remote sensing data set show that the designed network has advantages over the rest single-stage network, which has a better detection effect of the building area under the complex background.