自适应融合残差网在图像分类中应用研究

Research on the Application of Adaptive Fusion Residual Network in Image Classification

针对卷积神经网络存在随着网络深度增加导致优化困难,识别正确率降低、泛化性能差等问题,在Res Net(残差网络)基础上,提出了一种基于softmax全连接自适应门控网络融合模型.该方法在隐层网络深度达到一定层数后,设置多种卷积核尺寸作为独立网络输出,通过softmax全连接门控网络输出各模型选择概率,融合多种卷积尺寸残差网输出作为模型最终输出.实验表明,本文提出的融合残差网络模型更适合于多类别、精细化数据集,与单网络模型相比,在训练集上具有更好的收敛性,在测试集上具有更好的泛化性能.

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.