基于Segment Anything的堆叠球团粒度检测

Stacked pellet particle size detection based on Segment Anything

球团的合理粒度分布对其质量至关重要。通过深入研究现有方法,针对球团粒度检测中不完全分割和过度分割等问题,提出了Pellet-SAM(PelSAM)分割方法。该方法继承于Segment Anything(SAM)结构,主要操作是冻结图像编码器和提示编码器的相关参数,并在图像编码器上添加轻量级的空间和通道适配器及微调掩码解码器,使用标记的球团分割数据集对网络进行训练,从而实现对球团图像的准确分割。由于分割后的球团轮廓呈现类圆形,为了便于计算球团直径,先采用最小二乘法对分割后的球团轮廓进行圆拟合;再应用小孔成像原理,结合摄像头和球团表面的距离及摄像头的焦距等因素计算比例系数,然后利用该系数计算图像中球团颗粒的实际大小;最后根据球团颗粒的实际大小将球团粒度分布划分为4个范围,即小于8、[8,12]、(12,16]以及大于16 mm,其中[8,16] mm的为合格球。试验结果表明,该方法在球团分割数据集上取得了96.8%的准确率,比UNet和DeepLabv3方法分别提高了21.0和12.6个百分点。在实际工厂环境下,该方法计算得到的球团粒度分布与人工筛分测量的结果相近,最大误差仅为1.74个百分点。本研究为高效的球团粒度分析提供了可行的解决方案。

The reasonable distribution of pellet size is crucial for ensuring their quality.Through in-depth study of existing methods and addressing issues such as incomplete segmentation and over-segmentation of pellet size detec-tion,Pellet-SAM(PelSAM)segmentation method was proposed,which inherited the structure of Segment Any-thing(SAM),with key operations involving freezing relevant parameters of the image encoder and prompt encoder.It introduced lightweight spatial and channel adapters to the image encoder and fine-tuned the mask decoder.The network was trained using labeled dataset for pellet segmentation,enabling accurate image segmentation of pellets.Since the contour of the segmented pellet appeared almost circular,in order to make it easier to calculate the pellet diameter,the least squares method was used to perform circle fitting on the contour of the segmented pellet.Apply-ing the principles of pinhole imaging and considering factors such as the distance between the camera and pellet sur-face,as well as the camera's focal length,a scaling factor was computed.This factor was then used to calculate the actual size of pellet particles in the image.Finally,based on the actual size of pellet particles,the distribution of pel-lets was categorized into four ranges,less than 8,[8,12],(12,16],and greater than 16 mm,where pellets ran-ging from[8,16]mm were considered qualified.Experimental results show that this method achieves an accuracy of 96.8%on the pellet segmentation data set,which is 21.0 and 12.6 percent poin higher than the UNet and Deep-Labv3 methods respectively.In the actual factory environment,the pellet particle size distribution calculated by this method is similar to the result of manual screening measurement,with a maximum error of only 1.74 percent point.The research provides a feasible solution for efficient pellet size analysis.