An Hybrid Parallel Network Structure for Image Classification

Abstract

Abstract The image classification task is to divide certain image target into the specific class which they belong to. The deep neural network simplifies image classification by sending input image data into the neural network to learn feature extraction and having the output results automatically. Existing network structures have been able to achieve or exceed human capabilities in image classification tasks. However, these image classification technologies still have problems such as over-fitting network model, insufficient generalization abilities and long convergence time. Interference caused by complex environments may also cause the performance degradation. This paper proposes a hybrid parallel neural network structure for this kind of artificial intelligence task. By adopting the K-parallel connection CNN structure and using different scales of RNN/CNN structure mixed-layer modes, the effects of image’s size on different objects and different shooting angles can be reduced in image classification. Experiments on the datasets show that the proposed structure will have the advantage of accuracy and computational efficiency both in the validation datasets and the test sets.