Graph image representation from convolutional neural networks
A method for producing a graph representation of an input image, the method including the procedures of applying convolutional layers of a trained convolutional neural network on the input image, defining a receptive field of a last convolutional layer of the trained convolutional neural network as a vertex of the graph representation, defining a vector of a three dimensional output matrix of the last convolutional layer that is mapped to the receptive field as a descriptor for the vertex and determining an edge between a pair of vertices of the graph representation by applying an operator on a pair of descriptors respective of the pair of vertices.
1. A method for producing a graph representation of an input image, the method comprising the procedures of:
applying convolutional layers of a trained convolutional neural network on said input image;
defining a receptive field of a last convolutional layer of said trained convolutional neural network as a vertex of said graph representation;
defining a vector of a three dimensional output matrix of said last convolutional layer that is mapped to said receptive field as a descriptor for said vertex; and
determining an edge between a pair of vertices of said graph representation by applying an operator on a pair of descriptors respective of said pair of vertices.
2. The method according to claim 1 , further comprising a procedure of employing said graph representation of said input image for performing a visual task.
3. The method according to claim 1 , wherein said procedure of determining an edge further comprises a sub-procedure of scaling said edge according to a distance between said vertices of said pair of vertices.
4. The method according to claim 1 , wherein only a portion of said convolutional layers are applied on said input image.
5. The method according to claim 1 , further comprising a procedure of applying at least one fully connected layer of said trained convolutional neural network on said input image after said procedure of applying convolutional layers, and further comprising a procedure of backtracking an output of said at least one fully connected layer to said last convolutional layer.