According to the design requirements of a slag rake monitoring system for a certain power plant, machine vision technology is applied to design an algorithm for monitoring slag rake scrapers in dusty environments. The monitoring algorithm utilizes a lightweight YOLOv5s-SCB object detection model to detect abnormal states of the slag rake scrapers in the power plant. Due to the high dust levels in the environment where the slag rake operates, a DehazeFormer dehazing network is introduced at the front end of the YOLOv5s-SCB model and enhances by integrating scale, spatial, and channel attentions. Furthermore, to further enhance detection accuracy, the RAFT optical flow network is incorporated into the monitoring algorithm to extract motion characteristics of the scrapers. These motion features extracted by RAFT are fused with the convolutional features extracted by YOLOv5s-SCB. Finally, testing on 400 dust-laden images for scraper monitoring shows a false detection rate of 0% and a miss detection rate of 4.9%.