Industrial Smoke Images Generation Network Based on Bidirectional Recurrent Network

The monitoring of smoke emissions serves as a crucial reference for guiding operations in the steel industry. However, due to the scarcity of industrial smoke datasets, most deep learning methods encounter difficulties when addressing this problem. To address the issue of limited smoke data, this paper proposes a novel data augmentation method called the bidirectional recurrent generative network. This algorithm combines the concepts of recurrent generative networks and Laplacian pyramid networks to learn smoke generation from smoke-free blast furnaces and eliminate smoke from smoke blast furnaces. In order to overcome the limitation of recurrent generative networks in accurately depicting target contours, this paper introduces contour feature vectors, which provide precise characterization of the target contours through constraint-based feature vectors. Furthermore, to alleviate the difficulty of generation, the Laplacian pyramid model is introduced, which reduces parameters and learning time by learning from low-resolution images. To validate the effectiveness of the proposed algorithm, extensive ablation experiments and comparative trials are conducted on four different smoke image datasets from industrial scenarios with limited quantities. The results demonstrate significant improvements achieved by this paper’s algorithm. Compared to other generation methods, the proposed algorithm exhibits superior performance in smoke generation tasks, offering an effective solution to address the challenges of scarce industrial smoke data and the training difficulties faced by related deep learning algorithms.