Image based fruit category classification by 13-layer deep convolutional neural network and data augmentation

Abstract

Fruit category identification is important in factories, supermarkets, and other fields. Current computer vision systems used handcrafted features, and did not get good results. In this study, our team designed a 13-layer convolutional neural network (CNN). Three types of data augmentation method was used: image rotation, Gamma correction, and noise injection. We also compared max pooling with average pooling. The stochastic gradient descent with momentum was used to train the CNN with minibatch size of 128. The overall accuracy of our method is 94.94%, at least 5 percentage points higher than state-of-the-art approaches. We validated this 13-layer is the optimal structure. The GPU can achieve a 177× acceleration on training data, and a 175× acceleration on test data. We observed using data augmentation can increase the overall accuracy. Our method is effective in image-based fruit classification.