To make intelligent identification of leaked steam in complex industrial field, a temperature field characterization and recognition method using infrared vision technology is proposed. The steam leakage process is simulated to reveal the occurrence and development characteristics of its temperature field, including diffusion characteristics, hammer tail characteristics, dynamic characteristics, and centrality characteristics. The temperature layer of steam temperature field is extracted, the detailed characteristics of temperature distribution are analyzed, and the variable scale gray processing method is proposed to realize the high-definition image representation of steam infrared temperature field. To improve the identification speed and accuracy, the MASK-RCNN model is established to make deep learning and dynamic mining of hammer tail image features of leaked steam. In this way, the recognition accuracy of single hammer tail is about 90.71%, and that of the overall algorithm is up to 99.93%. The algorithm is tested with leaked steam recognition in power plant equipment operation. Results show that time consumed to process 5 consecutive frames is about 0.50 s, and steam leakage of various particle sizes can be identified quickly and accurately.