Two-Stage Recognition Algorithm for Untrimmed Converter Steelmaking Flame Video

In the process of converter steelmaking, identification of converter status is the basis of subsequent steelmaking control, directly affecting the cost and quality of steelmaking. Usually, the status of converter can be identified according to the flame of furnace port. In this paper, we propose a two-stage recognition algorithm to identify converter status using furnace flame video. In the first stage, we design a 2D feature extractor based on the shift module, aiming to capture temporal information in real time. In the second stage, an attention-based network is designed to get a more discriminative temporal attention, in which we model the distribution of temporal attention using conditional Variational Auto-Encoder (VAE) and a generative attention loss. We collect the video data and construct the corresponding data set from real steelmaking scene. Experimental results show that our algorithm can meet the accuracy and speed requirements.