부산대학교 도서관

Our study aims at developing an intelligentized coke size analysis system based on convolutional neural network for blast furnace ironmaking. For this purpose, a special image acquisition device was designed to capture high resolution images of cokes on a moving conveyor belt. A deep learning-based object detection technique was applied to extract the desired coke regions on the images. To improve the efficiency of detecting cokes with varying size, a set of anchor boxes were selected by an estimation function for the training procedure of an object detection model. According to experimental result, the proposed system could achieve recall and precision of 91.2% and 88.1 %, respectively. In addition, in the on-line validation, the analysis result by our system had an average difference of 2.9 mm (5.3%) by compared with the expert manual measurement.