Basic oxygen furnace endpoint carbon content prediction with Contourlet–ResNet representation of flame image features

Accurate prediction of the endpoint carbon content in basic oxygen furnace (BOF) steelmaking is critical for controlling steel quality. The carbon content oxidation rate in the molten pool is reflected in the textural changes of the furnace mouth flame. However, these textures exhibit complex, multi-scale, and multi-directional characteristics, presenting significant challenges for effective capture and analysis. Although convolutional neural networks have demonstrated excellence in image analysis, their varying sensitivities to different visual features result in limitations when capturing and representing the rich multi-scale and multi-directional texture information in flame images. To address these challenges, we propose a BOF endpoint carbon content prediction method with Contourlet–ResNet representation of flame image features. A dual-branch network architecture is constructed, leveraging Contourlet transform and ResNet to extract multi-scale, multi-directional texture, and deep semantic features. A Contourlet-guided ResNet feature optimization module is proposed, which adaptively weights deep features using Contourlet coefficients, enhancing the model’s ability to represent complex textures. Furthermore, a Contourlet coefficient global multi-directional feature extraction module is developed to effectively capture and fuse multi-directional global texture information at different scales of Contourlet coefficients, providing a complementary representation of texture information. The enhanced deep features are then fused with global multi-directional features to construct a comprehensive feature representation for endpoint carbon content prediction. Experimental results demonstrate that the proposed method achieves a prediction accuracy of 86.76% within a ±0.02% error range.