Abstract: A kind of convolution neutral network CNN-M was proposed to recognize the morphology of molten pool in CO2 welding. A simple pretreatment of the molten pool image was adopted as the input vector in CNN-M, which avoids the adverse effects brought by the subjectivity of artificial extraction of molten pool image features. In order to decrease the possibility of network over-fitting caused by the sparse training data set, several methods including ReLU activation function, random Dropout and SVM classifier were used in CNN-M. The experimental results show that CNN-M has better performance in recognition rate and recognition speed than the BP neural network does, which adopt the characteristic values of the molten pool as the input vector. The performance of CNN-M is able to meet the requirement of the on-line molten pool monitoring.