Traditional methods for mixed-particle classification usually extract particle features from binary images. After designing appropriate features according to the particle type, particles can be classified using widely known classifiers, such as back-propagation neural network and support vector machine (SVM). However, classifying touching particles is a challenging, and inappropriate feature design may further reduce the classification accuracy. Herein, a convolutional neural network (CNN) is utilized to extract the features for building mixed-particle image classifiers. In particular, particle locations in an image are determined using a region proposal network. Furthermore, a classifier is designed and combined with a fully convolutional network to achieve pixel-level particle segmentation. Experimental analysis is performed on some flowing-mixed-particle systems comprising spherical, elongated, and irregular particles. According to the analysis results, SVM method using manually designed features can achieve an average precision of 87% and recall of 87%, whereas those of the CNN-based method are up to 97% and 93%, respectively. The latter method can also reduce the analysis error by more than 11% for number median diameter (Dn50) of irregular particles. In addition, several shortcomings in traditional methods, such as the need for manually designed features are solved, making it easier to build an end-to-end system for effective real-time image analysis of flowing mixed particles.