Abstract: A side-scatter lidar is known to have evident advantages over other types of lidar in atmosphere detection. However, the signal of the side-scatter lidar may suffer from the noise as all other lidars. It is noted that the original signal of the side-scatter lidar is an image captured by a CCD camera. Therefore, denoising the side-scatter lidar signal may need more efforts than ordinary radar signals. In the paper, a denoising algorithm based on convolution neutral network is proposed for the side-scatter lidar signal. We combine the residual learning with batch standardization in the network. Further, attention mechanism and activation function in the network are optimized in order to improve the learning efficiency and the network output performance. Using the proposed algorithm, we successfully identify the noise and separate the noise from the simulated lidar signal. The signal-to-noise ratio is hence increased. Simulation results show that the peak signal-to-noise ratio is increased by over 5 dB using the proposed denoising algorithm. The relative error of signal is reduced to 9.62%. The proposed denoising algorithm based on the convolution neutral network is shown to be efficient for improving the side-scatter lidar signal, compared with the possible denoising algorithms based on wavelet transform and Wiener filtering.