Systems and methods for document image analysis with cardinal graph convolutional networks

A method for use in a document analysis system operable to process a document (102) in an improved manner based on a cardinal graph convolution network (CGCN), the method comprising the steps of:

• generating a cardinal graph representation (202) of the document, wherein in the cardinal graph representation each word in the document is represented as a single node on a graph and edges are connected between neighbouring nodes in four cardinal directions;
• generating a features tensor (F.T) (204) for each node of the cardinal graph representation, the features tensor having a dimension of n×m, where n is the number of nodes in the cardinal graph representation and m is the features vector size;
• encoding the cardinal directions in a 3×3 neighborhood to generate an adjacency tensor (A.T) (206) having node neighbour indices, the adjacency tensor (A.T) having a dimension of nl×3×3, where l is the number of layers in the (A.T) nl is the number of nodes in the layer;
• transforming the entries of the adjacency tensor (A.T) into a one-hot encoding of the node neighbour indices;
• creating a neighbourhood feature tensor (NFT) (208) through an Einstein-summation over the node indices of the 's (206) one-hot encoding and the node indices of the features tensor (F.T) (204) using the equation:
• NFTi,j,k,l=Σanli,j,k,aFTa,l,
• wherein neighbourhood feature tensor (NFT) having a dimension of nl×3×3×m where each 3×3×m block is a local patch of features in the neighbourhood of each node;
• scaling the features in each block by multiplying them with a connectivity weight tensor (C.W.T), wherein the connectivity weight tensor (C.W.T) encodes a weighting over the adjacency tensor (A.T) edges;
• convolving (210) the scaled features in each block with K learnable kernels of dimension 3×3×m in W(l) resulting in a convolved tensor of nl×3×3×K dimension;
• and
• reducing (212) the convolved tensor into a new feature tensor of dimension nl×K by selecting a center cell of the convolved tensor, wherein the center cell corresponds to a node's self-link in the adjacency tensor (A.T).