Three methods were compared to detect SiC particles from microstructure images of Al-SiC particle-dispersed composites. The software developed in our laboratory, MPImage, converts the microstructure image into a binary image and detects particles by contour extraction. In the remaining two methods, neural network algorithms called SSD (Single Shot Multibox Detector) and Mask R-CNN (Mask Region Convolutional Neural Network) were used to detect the particles. MPImage has the lowest undetected rate, but, it’s false detection rate was higher than the number of correct particles, suggesting that noises were detected as particles. In the SSD method, some particles were detected, but more than half were not detected. Mask R-CNN had the lowest false detection rate of 0% and the low undetected rate of 1.352%, when learning rate was adjusted. In addition, it can determine boundaries and detect contours even if the particles are adjacent to each other. Therefore, Mask R-CNN was the most suitable method for particle detection of particle-dispersed composites.