A convolutional neural network (CNN) was used to identify the morphology of rough particles from their interferometric images. The tested particles had the shapes of sticks, crosses, and dendrites as well as Y-like, L-like, and T-like shapes. A conversion of the interferometric images to polar coordinates enabled particle shape recognition despite the random orientations and random sizes of the particles. For the non-centrosymmetric particles (Y, L, and T), the CNN was not disturbed by the twin image problem, which would affect some classical reconstructions based on phase retrieval algorithms. A 100% recognition rate was obtained.