Investigation on mixed particle classification based on imaging processing with convolutional neural network

Abstract

For the classification of mixed particle with imaging method, the traditional processing techniques tend to extract the particle features from binary images, where particles can be classified by combining feature design corresponding to the particle type and classifiers, such as BP neural network and Support Vector Machine (SVM). However, the accuracies of these methods would be seriously influenced by particle agglomeration and imprecise feature design. In the work, the convolutional neural network (CNN) was introduced to extract the particle features in the image, where the particle localizations were obtained by implementing the region proposal network (RPN). Meanwhile, the particle segmentation at pixel level was achieved by a developed classifier and a fully convolutional network. A series of experiments were performed to a flowing mixed particle system consisting of spherical, elongated and irregular particles, and it showed that both the average accuracy and recall rate of SVM method were up to 87% with artificial feature design, while they were increased to 97% and 93% with CNN, respectively. For the median diameter (Dn50) of irregular particles, the latter method can also reduce the analysis error by more than 11%. It revealed that some shortages like imprecise feature design in traditional methods can be covered, and an end-to-end classified system can be formed to provide a more effective way for online analysis of flowing mixed particles.