Image Classification Based On Deep Convolutional Network And Gaussian Aggregate Encoding

With the rapid rise of deep learning, the deep convolutional neural network(CNN) has become a basic tool in image classification tasks. The depth and width of CNNs keep increasing, which leads to higher risk of overfitting and puts more requirements to the hardware. Therefore, how to improve the classification accuracy and reduce the parameters of the model is still a challenge. In this paper, we propose a novel image classification network based on deep CNNs and gaussian aggregate encoding, denoted as DCGAE. It combines existing CNNs and traditional encoding techniques to accomplish parameter reduction while retaining outstanding accuracy. In terms of the architecture, DCGAE first extracts image features by a deep CNN, which allows images of different sizes to be taken as input, and then append a gaussian aggregate encoding layer before classification. Extensive experiments conducted on multiple classification datasets demonstrate the effectiveness of our algorithm. Our model achieves state-of-the-art 98.30% accuary on STL-10 dataset.