Inc-ghostNet: A lightweight muti-branch ghost module fusion network for image classification

Convolutional neural network has been successfully applied in various domains, including image classification, target detection, and image segmentation. However, traditional model has the problem of large volume and low accuracy, which can not be used in mobile devices directly. To solve these problems, a lightweight multi-branch residual structure Network named Inc-ghostNet is proposed in this paper based on ResNet and Inception-Ghost module. Ghost module is used in Inc-ghostNet to extract features for the high recognition and reduced model parameters. DFC(decoupled fully connected) attention mechanism is used to gather channel information of the network to improve it accuracy. Moreover, Mish activation function is used to improve the non-linear expression ability of the network. The performance of the proposed network is studied using image recognition database. Experimental results show that compared to ResNet50 network, the recognition accuracy of the proposed Inc-ghostNet is improved by 2.1% , while the number of model parameters is reduced by 87.1%, the volume of model is reduced by 89.7%, and the calculation amount is reduced by 91.9%, respectively. In addition, Inc-ghostNet also shows more excellent performance than other compared lightweight networks for different image datasets.