This paper presents a single-image rain removal algorithm utilizing generative adversarial networks with dilated residuals and patch constraints, addressing the limitations of previous image rain removal methods, such as the loss of detail in rain-removed images and the introduction of artifacts. Firstly, the generator is structured as a multi-level dilation residual network to effectively eliminate noise signals from the input image while extracting features associated with rain streaks, thereby facilitating both deconstruction and reconstruction of high-frequency information. Secondly, a patch discriminator is incorporated to establish a patch-level adversarial framework that constrains the generator, enabling more flexible and efficient artifact discrimination by the model. Finally, a concept of spatial averaging is integrated into the loss function to enhance training efficacy for the network model, further improving its ability to retain details. Experimental results demonstrate that our method performs well in deraining both actual rainfall images and Rain800 dataset while exhibiting strong predictive capability for distorted regions within images, thus ensuring effective restoration of content details.