Dynamic Graph-Driven Heat Diffusion: Enhancing Industrial Semantic Segmentation

Dust significantly impacts construction progress and worker health, necessitating the use of machine learning for dust area identification and pollution mitigation. Existing dust semantic segmentation methods face limitations due to data quality, leading to suboptimal performance in real-world applications. These limitations stem from complex backgrounds and the model's inadequate ability to extract dust features, resulting in conflicting training processes. To overcome these challenges, we propose a thermal energy propagation training approach. Our method incorporates a thermal propagation mechanism into the self-training framework to dynamically adjust the weights of the training loss. This dynamic weight assignment is achieved through a graph-based approach. By utilizing ground truth, we generate an initial weight distribution that gradually weakens from the center to the surrounding areas, expanding as training progresses. This dynamic weighting scheme enables the model to gain insights into pixel clusters in the early stages and focus on learning edge areas in later stages. Throughout training iterations, the teacher model guides the dynamic adjustment of weight distribution, continuously updating it to capture essential regional information. This dynamic graph-driven weight assignment enhances the model's ability to accurately extract dust features. Experimental evaluations conducted on diverse datasets demonstrate the competitive performance of our approach compared to previous methods. The proposed dynamic graph-driven heat diffusion technique addresses the limitations posed by complex backgrounds and data quality, making it a valuable tool for dust area identification. It enables improved construction progress monitoring and worker health protection in industrial settings.