In the environment with small Radar-Cross Section (RCS), serious clutter and noise, it is difficult to detect the slow-moving target at short range. To solve the problem, a target detection method based on double-channel convolutional auto-encoder with skip connection is proposed. The time-frequency spectrum is introduced to the convolutional auto-encoder as the input. The neural network structure adopts the IQ double-channel to extract amplitude features and phase features from target echoes, and fuses the multi-dimensional features in the middle layer. Considering that the target scale is small in time-frequency spectrum, the skip connection structure is designed in the network, which connects the top and the bottom of the network to improve the recovery ability of target feature in decoders. Moreover, it can mitigate the gradient dissipation problem of the deep network, and improve the efficiency of end-to-end training. Experimental results show that: In the environment with serious clutter and noise, this method can achieve better performance on detection of small, slow-moving target than the traditional methods.