Target Extraction of Carbonaceous Substances in Blast Furnace Dust Based on Pyramid Scene Parsing Network

This paper proposes an improved pyramid scene parsing network (PSPNet) to accurately extract carbonaceous substances from blast furnace dust microscopic images, which addresses the challenges posed by multiple components, blurred edges and similar structures among different categories in these images. Based on the PSPNet architecture, the proposed network selects the lightweight network MobileNetV2 as the backbone feature extraction network to extract features from the input images; and introduces a Receptive Field Block (RFB) module to enhance the deep feature from MobileNetV2; Finally, feature maps enhanced with RFB are embedded into the Pyramid Pooling module and fused with the previous feature maps to obtain the final segmentation results. According to the classification criteria, a semantic segmentation dataset with seven categories of blast furnace dust microscopic image is constructed, and the experimental results on this dataset show that, by using the improved PSPNet model proposed in this paper, the recognition of carbonaceous substances targets reaches 86.98% of the mIoU, and the mPA is 93.78%, which are 9.1% and 12.45% higher than those of the original model, respectively.