Smoke recognition in steelmaking converter images: Class-distance-based feature selection model

Controlling the smoke and dust generated in steelmaking efficiently and precisely is a significant task for green metallurgy. This study aims at recognizing the smoke and dust generated from converters in steelmaking processes, to reach this goal, a research about feature generation and selection is conducted. The comprehensive feature datasets include GLCM features, LBP features and gray statistics features as a feature vector of 212 dimension for each image. To obtain better classification accuracy and performance, a supervised Class-Distance-based Feature Selection (CDFS) algorithm is proposed which involved the Mahalanobis distance which is chosen as class distance, correlation, T-test and train with machine learning classifier SVM. Then classifying images as smoke or non-smoke, the best feature combination with high classification performance as well as smoke images classification model can be exported as an output. Applied results on 5 converter image datasets show that this feature selection algorithm after optimization of Z-Score data standardization can not only reach an excellent accuracy in range of 92.54%-99.18% and AUC from 0.88 to 0.99 with feature size of 6–12 dimension, but apply well on small sample and imbalance class datasets. Compared with PSO, GA and BA, CDFS obtain higher accuracies as well as lower feature size. An application result also proved the effectiveness and generalization of this algorithm which means it can achieve precise classification accuracies with less cost in training as well. It can provide the foundation for efficient and precise emission treatment of converters in steelmaking.