In this paper, a deep convolutional neural network (CNN) is used to remove speckle noise from synthetic aperture radar (SAR) images. However, only applying CNN to remove noise causes an under-fitting problem. To overcome this issue, we suggest to use stationary wavelet transform (SWT) to the images as a pre-processing. Afterward, the resultant sub-band images are utilized to construct the similar sub-band images to the original images by training the CNNs. The training process is carried out by considering a large multi-temporal SAR image and its multi-look version. In the experiment result of this paper, the proposed method showed better performance compared to other denoising algorithms in regard to PSNR and SSIM.