Balling phenomenon is the most common defect in selective laser melting(SLM) forming process,which affects the fatigue life and physical properties of the final formed part. Reasonable control of the occurrence of balling phenomenon is of great significance to maintain the continuous forming process and obtain high quality formed parts. In this paper,a balling degree characterization method was proposed based on the study of balling feature extraction method for parts surface during SLM forming process. The effectiveness of the method through orthogonal experiments is verified,and the correlation between balling degree and laser power density was also established. Moreover,the grade of balling degree was defined,and finally a deep convolutional neural network(CNN) model was constructed to automatically identify the grade of balling degree of the part surface to assist the experiment and production personnel to make corresponding decisions. The recognition results showed that the network recognition accuracy reaches 96.4% on the small segmentation image set,and the recognition accuracy reaches 100% on the global microscopic image set,which achieved a good recognition effect. This paper will provide an effective way to realize real time control of forming quality in the SLM forming process.