Abstract: Aiming at the problems of traditional molten pool detection methods in arc welding additive manufacturing process, such as relying on empirical parameters, low accuracy and long recognition time, a molten pool shape detection method based on additional loss function of edge included angle is proposed to realize rapid and accurate detection and recognition of molten pool. Firstly, the molten pool is represented on multiple scales by integrating multiple features with feature pyramid network, and the dependence on empirical parameters is avoided. Secondly, PointRend neural network module is used to optimize the sampling points based on fine-grained features and coarse prediction mask to reduce the time required for molten pool target detection and recognition. Thirdly, the additional loss function of edge included angle is studied to maximize the classification interval in the angle space, which makes the features extracted by the network more separable and improves the recognition accuracy of model. Finally, tests are carried out using the actual molten pool monitoring data, and the results show that the proposed method has a high recognition accuracy, the accuracy rate is 97.85%. In the presence of molten droplet coverage interference, it can also achieve an accurate detection and recognition performance. Compared with the detection width and measured width of molten pool, the absolute error is less than 0.36 mm. The experimental results have demonstrated that the method is effective and reliable.