To address the issue of low accuracy in diagnosing irregular defects under multisource noise, a multiscale convolutional attention neural network (CANN) magnetic flux leakage detection method was proposed. First, the improved graph convolutional neural network was proposed to extract the long-range correlation between pixel-level and structural-level features of the leakage magnetic signal, and noise features were introduced to reconstruct the leakage magnetic signal to reduce the complex fluctuations caused by multi-source noise. Then, the decoupling fully connected attention mechanism was introduced into the traditional convolutional neural network (CNN) to aggregate the central features of the reconstructed signal and further extract the tiny features of irregular defects. Finally, the mean square error was used to determine the threshold value to detect irregular defects on the pipe surface. The results show that the average detection accuracy of the proposed method for irregular defects reaches 96.7%, which is 5.5% higher than the best detection accuracy of the existing methods.