【摘要】 传统卷积神经网络对空间信息不敏感,无法学习到不同特征间相对位置的关系,且每一层神经元的感受野被设计为相同大小,导致提取的图像特征信息不够精确。针对这些问题,提出一种选择性卷积核胶囊网络用于图像分类任务。在经典胶囊网络的卷积层融入具有两个分支的选择性卷积核网络,以提取更为丰富、准确的图像特征信息,提高图像分类准确率。采用CIFAR-10、Fashion-MNIST、SVHN经典图像分类数据集进行实验,结果表明,相比于基线胶囊网络模型,新模型的识别精度更高,尤其在CIFAR-10数据集上识别精度提高了1.73%,从而有效提升了图像分类准确率,具有良好的图像识别能力。更多还原

【Abstract】 The traditional convolution neural network is insensitive to spatial information and cannot learn the relative position relationship between different features,and the sensory field of each layer of neurons is designed to be the same size,which leads to the inaccuracy of the extracted image feature information. To address these problems,a selective convolutional kernel capsule network is proposed for image classification tasks. The convolution layer of the classical capsule network is integrated into the selective convolution kernel network with two branches,which can extract more abundant and accurate data image feature information and improve the accuracy of image classification. The experimental use CIFAR-10,Fashion-MNIST,SVHN these classical image classification data sets.The results show that the recognition accuracy of the new model is higher than that of the baseline capsule network model,especially the recognition accuracy on the CIFAR-10 data set is improved by 1.73%. The new model effectively improves the accuracy of image classification and has good image recognition ability.更多还原