Due to the influence of external environment, LiDAR images are easily affected by various noises, which reduces the accuracy of data. To this end, a sparse denoising method for LiDAR images based on deep learning is proposed. We use an accelerated backward projection algorithm to generate initial LiDAR images. In response to the image blurring phenomenon generated during the imaging process, we set adaptive transition points and enhance blurry contrast to complete the deblurring processing of the LiDAR images. Combining the advantages of deep learning technology, an adaptive stack style sparse denoising autoencoder is established. Through multi-channel SRDA, each SDA is trained for different types of noise, and finally linearly combined to handle multiple types of noise simultaneously. This multi-channel approach can more comprehensively eliminate various noises and improve the sparse noise reduction effect of LiDAR images. The experimental results show that the proposed method not only effectively removes the blurring phenomenon of LiDAR images, but also has a relatively efficient denoising ability.