A general multi-scale image classification based on shared conversion matrix routing

Yuxiao Wang, Kai Li, Yu Lei

Abstract: For the deep convolutional neural network, the input image needs to be fixed to the corresponding size due to the existence of the fully connected layer. Stretching and clipping can make the image reach the required size, but these operations can easily distort the image. The methods based on pooling layer groups enable variable-size feature maps to be converted into fixed-size. However, there is a loss of information due to the pooling operations, and the recognition accuracy will be significantly reduced. Based on this problem, we propose the shared conversion matrix routing (SCMR) layer as a general network layer to replace the fully connected layer of the convolutional neural network, which can enable the network added to this layer to deal with multi-scale image problems without changing the original convolution structure and parameters. In the SCMR layer, we propose a RECOMBINATION method which dynamically increases or decreases the number of capsules according to the scale of the input image to ensure the normal operation of the convolutional layer and the SCMR layer. At the same time, a new dynamic routing algorithm is established by sharing the transformation matrix in the SCMR layer so that the SCMR layer can receive the convolutional multi-dimensional feature map and generate fixed-size image features output to realize the classification of multi-scale images. The algorithm makes each capsule have a corresponding weight to avoid the problem of feature loss, which improves the recognition rate. In addition, new capsules are created by increasing the dimensions of capsules to solve the exploding gradient problem in backpropagation. The experimental results show that the accuracy of the method proposed in this paper is better than the modern methods on public datasets.