基于改进Multi-ResUnet算法的黏结集料图像分割方法

Image Segmentation Method for Cohesive Aggregates Based on Improved Multi-ResUnet Algorithm

摘要:

为提高黏结集料图像的分割精度，提出了基于Inception网络与残差连接优化的黏结集料图像分割模型（Multi-ResUnet模型）。利用实验室自主研发的集料三维特性分析系统V3.0对黏结集料图像进行采集，并建立图像分割模型样本集，然后采用图像分割模型对样本集进行训练。结果表明：相较于分水岭算法和Unet模型，该图像分割模型的精确率分别提升了30.46%和2.11%，召回率分别提升了4.68%和1.85%，准确率分别提升了25.95%和2.47%。

Abstract:

In order to improve the image segmentation accuracy for cohesive aggregates， a image segmentation model （Multi-ResUnet model）based on Inception network and residual connection optimization was proposed. The three-dimensional aggregate characterization system V3.0 developed by the laboratory itself was used to collect the images of cohesive aggregates and the sample set of the image segmentation model was created. Then， the image segmentation model was used to train the sample set. It is shown that compared with the Watershed algorithm and Unet model，the image segmentation model improves 30.46% and 2.11% in precision， 4.68% and 1.85% in recall， and 25.95% and 2.47% in accuracy， respectively.