The invention discloses a deep learning based automatic defect identification method for an underground pipeline. Positive and negative sample sets, of the underground pipeline, needed by training ofconvolutional neural network (CNN) are prepared; the sample sets are preprocessed, and modified, in a batch manner, to the uniform size of 300*300, data improvement is carried out, and sample data fortraining is generated; a structure of the CNN is designed, training is carried out, and a weight connection matrix W during network convergence is obtained and used for a detection process later; aimed at video data, first and last 10 frames of a video are eliminated, a defect target frame is selected roughly, and key frames are sampled from the video every 10ms; each sampling frame of the videois input into the CNN, and whether there is a defect is determined; and according to a result of each frame in the last step, whether the video includes defects is concluded. According to the method,the utilization rate of data is improved, characteristics of a defect pipeline image are learned automatically via the convolutional network, and automatic identification for the defect pipeline is realized.