基于深度学习的粗骨料在线检测分割方法研究

Research on Online Detection and Segmentation Method of Coarse Aggregate Based on Deep Learning

使用图像法检测骨料粒径时,图像分割的质量是影响骨料粒径检测的重要因素。目前,骨料图像分割已经从传统的分水岭算法和阈值分割算法发展到使用实例分割算法对堆叠的粗骨料进行分割。针对ISTR模型未分割骨料较多的问题,提出了一种改进算法和评估网络模型的评价指标,对优化前后的网络模型进行对比实验。实验结果表明:与原网络模型相比,优化后算法MIoU提升了3.4%,达到82.6%;未分割的骨料占比降低了8.2%,达到9.4%;检测分割能力提升明显,证明所提方法在骨料检测分割任务中的可行性与有效性。

When using the image method to detect aggregate particle size,image segmentation quality is an essential factor affecting the detection of aggregate particle size.At present,aggregate image segmentation has developed from the traditional watershed and threshold segmentation algorithms to the instance segmentation algorithms to segment stacked coarse aggregates.Aiming at the problem of many undivided aggregates in the ISTR(end-to-end instance segmentation with transformers)network model,an improved algorithm and an evaluation index were proposed for evaluating the network model.Finally,a comparative experiment was conducted on the network model before and after optimization.The experimental results show that compared with the original network model,the MIoU(Mean Intersection over Union)of the optimized algorithm has increased by 3.4%,reaching 82.6%,the proportion of unsegmented aggregates has decreased by 8.2%,reaching 9.4%,the detection and segmentation ability improves significantly,which proves that the feasibility and effectiveness of the proposed method in aggregate detection and segmentation tasks.