Connectivity similarity based graph learning for interactive multi-label image segmentation
View Patent ↗A system and method of connectivity-based image processing to identify and extract objects in image data is described. Variations on the method may include iterative local smoothing operations and various algorithmic solutions to improve real-time processing. Variations may also include object extraction processes based on user-provided information about an object in an image.
1. A method of providing image data, the method comprising:
identifying information associated with an object in an image;
constructing a graph based on the identified information associated with the object in the image, wherein the graph includes connectivity similarity for a plurality of nodes in the graph and wherein constructing a graph includes:
defining effective distances for valid paths between two graph nodes and choosing the minimum value of the effective distances as an inter-nodal distance between the two graph nodes;
labeling a graph node with a number of labels corresponding to the number of nodes to which the graph node is connected, wherein a label of the node is associated with a connectivity distance between the graph node and a connected node; and
generating a connectivity distance matrix, wherein the matrix has at least one column vector that corresponds to a graph node's confidence for a specific label;
performing local smoothing between spatially close graph nodes, wherein performing local smoothing between spatially close graph nodes comprises performing an iterative method that includes:
detecting an overall noise level in the matrix;
removing an outlier from the at least one column vector, wherein the outlier is determined based on mean confidence values of neighboring nodes; and
if the overall noise level is above a specified threshold, repeating the detecting and the removing until the noise level is at or below the specified threshold; and
providing image data representing the object based on results of the local smoothing.
2. The method of claim 1 , wherein constructing a graph includes establishing seed nodes based on the identified information.
3. The method of claim 1 , wherein identifying information includes receiving the information associated with an object in an image.
4. The method of claim 1 , wherein performing local smoothing between spatially close graph nodes includes executing an iterative linear neighborhood propagation algorithm.
5. The method of claim 1 , wherein at least one graph node has more than two labels.
6. A non-transitory computer-readable medium having embodied thereon a set of instructions that, when executed, causes a computer to perform a method of providing image data, the method comprising:
identifying information associated with an object in an image;
constructing a graph based on the identified information associated with the object in the image, wherein the graph includes connectivity similarity for a plurality of nodes in the graph and wherein constructing a graph includes:
defining effective distances for valid paths between two graph nodes and choosing the minimum value of the effective distances as an inter-nodal distance between the two graph nodes;
labeling a graph node with a number of labels corresponding to the number of nodes to which the graph node is connected, wherein a label of the node is associated with a connectivity distance between the graph node and a connected node; and
generating a connectivity distance matrix, wherein the matrix has at least one column vector that corresponds to a graph node's confidence for a specific label;
performing local smoothing between spatially close graph nodes, wherein performing local smoothing between spatially close graph nodes comprises performing an iterative method that includes:
detecting an overall noise level in the matrix;
removing an outlier from the at least one column vector, wherein the outlier is determined based on mean confidence values of neighboring nodes; and
if the overall noise level is above a specified threshold, repeating the detecting and the removing until the noise level is at or below the specified threshold; and
providing image data representing the object based on results of the local smoothing.
7. The non-transitory computer-readable medium of claim 6 wherein identifying information includes receiving the information associated with an object in an image.
8. The non-transitory computer-readable medium of claim 6 wherein performing local smoothing between spatially close graph nodes includes executing an iterative linear neighborhood propagation algorithm.
9. The non-transitory computer-readable medium of claim 6 , wherein at least one graph node has more than two labels.
10. The non-transitory computer-readable medium of claim 6 , wherein constructing a graph includes establishing seed nodes based on the identified information.