基于迁移学习和图像超分的紧密堆积粗集料级配的实时分析

Effective Real-time Low-quality Image Analysis on Stacked Coarse Aggregate Characteristics by Transit Learning and Super Resolution

针对在复杂生产环境中获取图像质量受限的情况,提出使用迁移学习方法的分割模型帮助生产环境模型训练收敛,以及使用图像超分方法增强图像质量和扩大粗集料目标面积等改进策略,对生产场景下质量受限的粗集料图像进行有效分割和分析。结果表明,所得到的分割模型Model-L,不仅相比于在生产环境下直接训练的模型Model-P的mAP值提高了0.3以上,EMS值提高了27%,且在不同粒径范围的粗集料区间的有效分割数量均远远优于Model-P,级配结果也较为可靠,5~25 mm范围内与机械初筛的误差在10%以内。所提出的策略展现了一定的普适性,对于部分网络粗集料图片也具有良好的分割性能。

This study addressed the challenges of limited image quality in complex production environments by proposing the use of transfer learning methods to aid in the convergence of models trained in production environments.The study also introduced strategies such as image super-resolution to enhance image quality and expand the target area of coarse aggregates,enabling effective segmentation and analysis of coarse aggregate images under quality constraints in production scenarios.The results demonstrate the effectiveness of the proposed strategies.The segmentation model,Model-L,not only improves the mAP value by more than 0.3 and increases the EMS value by 27% compared to Model-P trained directly in production environments,but also excels in the segmentation quantity of coarse aggregates in various particle size ranges.The gradation results are reliable,with errors within 10% of mechanical sieving in the 5~25 mm range.The strategies proposed in this study exhibit a certain level of generality and demonstrate good segmentation performance for coarse aggregate images in specific network scenarios.