Abstract: In order to accurately recognize the weld target in complex environment, a visual model of welding robot based on deep learning was established. The model adoped a convolutional neural network (CNN) combining local connection and full connection. The local connection was composed of 3 convolution layers (C) and 3 subsampling layers (S) with C-S alternating mode for feature extraction of welding target. The full connection layer was composed of input layer, hidden layer and output layer as a classifier for weld target recognition. More than 1 000 image samples of welding targets were sampled for CNN network training, and the influence of different CNN structure parameters on the model was analyzed. The test results show that the visual model was robust to the translation, rotation and scaling of welding targets, and could be applied to the visual navigation of welding robots.