In order to reduce the interference of noise signal and device performance in the process of transmission line tower defect detection, and improve the accuracy and efficiency of transmission line tower defect detection. A defect detection method of transmission line towers based on Generative Adversarial Network (GAN) and Improved Region Proposal Network (RPN) is proposed. GAN is used to collect the salient images of transmission line towers, and the semi-soft threshold function model is used to eliminate the noise in the images, so as to avoid the influence of noise on the defect detection process. The contour features of transmission line tower image are extracted by random forest decision tree, and RPN is improved based on multi-scale algorithm. The features are input into the improved RPN model, and the defect detection of transmission line tower is completed by defect location and segmentation. The test results show that the proposed method has high accuracy and good defect detection effect and efficiency, which is beneficial to improve the quality of transmission line tower defect detection and reduce the occurrence of power accidents.