Class Prediction Performance Boost Using Gaussian Mixture Models

Fine-grained image classification remains a challenging task due to subtle visual differences between subordinate categories. Most existing approaches focus on improving key parts detection ability and feature extraction ability of neural networks, while our approach focuses on improving class prediction ability with Gaussian mixture models. Gaussian mixture layers (GML), which learn Gaussian mixture feature distribution and predict classes, are designed to replace fully connection layer, or placed right after them. Experiment results indicate that classification performance can be improved significantly by using GML on top of VGG model. In addition, GML are compatible with any neural networks that contains fully connection layers, which enables them to further boost the performance of cutting-edge classification networks.