混合颗粒系重叠图像分割与分类方法研究

针对传统图像处理算法对重叠颗粒的分割困难,引入Mask R-CNN深度学习算法并做针对性改进,通过调整残差网络ResNet-101加速训练,提出双FPN结构实现全局特征融合,使用Soft-NMS避免重叠颗粒漏检。设计了颗粒重叠图像实验系统,采集单一球形、球形与不规则混合多分散颗粒重叠图像用于分割研究。实验结果表明:该方法对混合颗粒分类准确率为91%,召回率为92%,均优于传统算法;其应用于含气泡的一水柠檬酸结晶过程中结晶的在线测量,所得结晶颗粒中位径误差为3.8%,数目误差为-1.3%。所提方法为混合颗粒的重叠图像分析提供了思路,后续有望解决图像法结晶过程后期在线监测乏力与气泡干扰的问题。

Research on Segmentation and Classification Methods of Mixed Overlapped Particle Images

Mask R-CNN was introduced to overcome the segmentation difficulties of traditional image processing algorithms for overlapped particle images.By adjusting residual network ResNet-101 to accelerate training, a double FPN structure was proposed to achieve global feature fusion, and soft-NMS was used to avoid o overlapped particle missing detection.A particle overlapped image experiment system was designed to acquire single spherical, spherical and irregular mixed multi-dispersed overlapped particle images for segmentation analysis.The experimental results show that the present classification accuracy is 91%, and the recall rate is 92%, which are both better than the traditional algorithms.When applied to the real-time measurement of crystallization and bubbles in the crystallization process of citric acid monohydrate, the method yields the errors around 3.8% for median diameter and -1.6% for the counting number of crystal particles.The proposed method provides a clue for analysis of overlapped mixed particle images, which is expected to solve the problems of image analysis at late stage of the crystallization process and eliminate the interference of bubbles involved during real-time monitoring.