In order to improve PET image quality, a Dilated U-Net(D-Unet) based PET image reconstruction method was proposed. First, to better enhance the contextual semantic information and extract deeper image features without increasing the number of parameters, we designed the residual density blocks nested with dilations(RnD Blocks) to process the PET image with high noise and striatal artifacts. In addition, the end-to-end PET image reconstruction network was constructed by combining the perception loss based on the pre-trained VGG network feature, rather than traditional mean square error loss(MSE loss), as the training loss function to retain the image details. The experimental results show that this method reduced complexity and kept a high convergence speed, and at the same time, it better suppressed noise. Compared with traditional method, the proposed algorithm is significantly improved in reconstruction effect.