In the cement production process, accurate monitoring of dust is a core link in building a solid line of defense for safe production and strictly controlling environmental pollution. Addressing the gap of lacking high-quality, large-scale annotated datasets in current cement dust semantic segmentation research, this paper constructs and releases a cement dust semantic segmentation dataset containing 100,000 high-resolution images. All images have undergone precise pixel-level annotation, which can provide reliable supervision signals for model training. On this basis, this paper systematically evaluates eight mainstream semantic segmentation models and presents comprehensive performance benchmarks and comparative analysis. This research provides important data support and evaluation benchmarks for visual inspection and intelligent monitoring in cement scenarios, and is expected to boost the continuous optimization and development of cement dust segmentation algorithms.