Parallel concatenated network with cross-layer connections for image recognition

Abstract

The traditional convolutional neural networks are heavy with millions of parameters and the classification accuracy is not high. To address this issue, we propose a novel model called parallel concatenated convolutional neural network with cross-layer connections. The model mainly includes parallel processing and concatenate operation. In parallel processing, the diversity of features is increased by using different sizes of convolution kernels. The parallel outputs are integrated together by concatenate operation. Meanwhile, an improved cross-layer connection structure is also added to the model. At the experimental stage, the model was tested on the Caltech-256 and Food-101 datasets, the experiment results indicate that the constructed PCNet (without cross-layer connections) increases the recognition accuracy by 2.54% and 7.31% compared to AlexNet, and the proposed RPCNet (with cross-layer connections) is improved by 6.12% and 12.28% compared to AlexNet.