Abstract: In robot automatic welding, accurate and efficient weld identification is the key to achieve high quality welding. Aiming at the problems of low efficiency and poor accuracy of the existing visual inspection methods, a welding seam recognition method based on full convolutional neural network was proposed. Firstly, the method collects data to train the full convolutional neural network and obtains the best network parameters. Then, the trained fully convolutional neural network and the best network parameters were used to semantically segment the weld images, and the weld area was separated from the background. Then, the skeleton of the segmented weld area was extracted to obtain the weld with a width close to a single pixel. Then, the shape of the weld is determined according to the self-defined straightness parameters to determine whether the weld is a straight line, and the least square method is used to carry out the line or curve fitting to get the final weld trajectory. The experimental results show that the proposed method can quickly and accurately identify the weld position and shape, which can be used as the technical basis for the automatic welding robot trajectory planning and control.