To enhance the accuracy and efficiency of defect detection in transmission line towers by reducing interference from noise signals and device performance, a defect detection method utilizing generative adversarial networks (GAN) and an improved region proposal network (RPN) was proposed. GAN was employed to capture significant images of transmission line towers, while a semi-soft thresholding function model was utilized to remove noise from the images and mitigate its impact on the defect detection process. The contour features of transmission line tower images were extracted using a random forest decision tree, and an enhanced RPN based on a multiscale algorithm was introduced. By inputting these features into the improved RPN model, defect localization and segmentation were performed for accurate defect detection in transmission line towers. Experimental results demonstrate the high accuracy, effectiveness, and efficiency of the proposed approach, thereby contributing to better quality control and a reduced occurrence of power accidents in transmission line towers.