Research on the Application of Adaptive Fusion Residual Network in Image Classification
YANG Jing-dong1，YANG Xin1，ZHAO Cheng2
1（Autonomous Robotics Laboratory，School of Optical-Electrical and Computer Engineering，University of Shanghai for Science and Technology，Shanghai 200093，China）2（Shanghai Integrative Medicine Hospital Affiliated to Shanghai University of TCM，Shanghai Institute of Vascular Disease of Integrative Medicine，Shanghai 200082，China）
Abstract：For the convolutional neural network，the increase of depth of network results in more difficult optimization，lower recognition accuracy，and poorer generalization performance.With respect to ResNet（residual network），we propose a self-adaptive fusion model based on softmax gating fully connected network.When the layers go beyond certain numbers，multiple convolution kernel sizes are used as the independent outputs respectively.According to the softmax gating fully connected network，we select the output probabilities for multiple models and combine ResNets with multiple convolution kernel sizes to acquire the final outputs.The experiments show that the fusion model proposed in this paper is more suitable for multi-category and refined datasets.Compared with the single ResNet model，the fusion model has better convergence performance for training datasets and better generalization performance for testing datasets.
杨晶东,杨鑫,赵诚. 自适应融合残差网在图像分类中应用研究[J]. 小型微型计算机系统, 2020, 41(2): 399-405.
YANG Jing-dong,YANG Xin,ZHAO Cheng. Research on the Application of Adaptive Fusion Residual Network in Image Classification. Journal of Chinese Computer Systems, 2020, 41(2): 399-405.
2020, Vol. 41 
