This study uses an end-to-end method for image deblurring based on deep learning to encode the blurred image and to subsequently decode it into a high-definition image. However, the lack of extracted feature information during encoding decreases the quality of the reconstructed deblurred image. To solve this problem, we propose two methods for improving the network structure. First, a dense network structure is added to the autoencoder network for extracting considerable feature information. Second, a multiscale perceptual field structure is introduced to extract considerable contextual feature information, comprising 4 scales of average pooling layers and up-sampling layers. The two improved methods achieve good image deblurring effects using the GOPRO and Kohler datasets.