Patent
Connection information normalization system based on adjacent matrix, graph feature extraction system and graph classification system and method
Reads0
Chats0
TLDR
In this paper, a connection information normalization system based on an adjacent matrix, a graph feature extraction system and a graph classification system and method was proposed, where the connection information elements in the adjacent matrix corresponding to the graph were concentrated into a specific diagonal region of the adjacent matrices, and the non-connected information elements were reduced in advance, so that when the window with the fixed size is used for traversing along the diagonal region, the sub-graph structures of all the corresponding sizes in the graph can be captured, and time complexity is greatly reduced.Abstract:
The invention provides a connection information normalization system based on an adjacent matrix, a graph feature extraction system and a graph classification system and method. The connection information elements in the adjacent matrix corresponding to the graph are concentrated into the specific diagonal region of the adjacent matrix, the non-connected information elements are reduced in advance, so that when the window with the fixed size is used for traversing along the diagonal region, the sub-graph structures of all the corresponding sizes in the graph can be captured, and time complexity is greatly reduced; further, a sub-graph structure of the graph is extracted along the diagonal direction by using the filtering matrix, and then the laminated convolutional neural network is adopted to extract a larger sub-graph structure. The invention is advantageous in that on one hand, the calculation complexity and the calculation amount are greatly reduced, and the limitation of calculation complexity and the limitation of window size are solved; the sub-graph structure of the large multi-vertex can be captured through a small window, and the depth feature of the implicit correlationstructure from the vertex and the edge is obtained, and the accuracy and the speed of image classification are improved.read more
Citations
More filters
Patent
Compound property analysis method and device, equipment and storage medium
TL;DR: In this article, the authors proposed a method to acquire a feature vector of a compound according to the molecular structure of the compound, wherein the feature vector comprises a node vector of each node and an edge vector of an edge of each edge.
Patent
Convolutional neural unit block, formed neural network and image classification method
TL;DR: In this article, a convolutional neural unit block is formed by stacking n convolution kernels in different directions and an m * m convolution kernel and then connecting with a of 1 * 1 convolutions kernels, wherein a is equal to the number of channels of the input feature map, and the block also comprises hop connection of identical transformation.
Patent
Image processing system and method based on deep learning, terminal and medium
TL;DR: In this paper, an image processing system and method based on deep learning, a terminal and a medium is presented. And a brand new convolutional neural network for line segment detection in an image is provided especially for the thrust capability of global semantics of the image.