System and method for performing saliency detection using deep active contours
A system and method are provided for performing saliency detection on an image or video. The method includes training and creating deep features using deep neural networks, such that an input image is transformed into a plurality of regions, which minimizes intra-class variance, and maximizes inter-class variance, according to one or more active contour energy constraints. The method also includes providing and output associated with the deep features.
1. A method of detecting object boundaries, the method comprising:
obtaining an input image;
processing the input image using a series of input layers in a deep neural network to obtain a set of global features;
using outputs of a plurality of the input layers to perform edge detection at different scales and combining the outputs to obtain a set of local features;
combining the global and local features and applying an optimization function that has been trained to minimize intra-class variance and maximize inter-class variance according to one or more active contour energy constraints, to enhance at least one object boundary from the input image; and
providing an output comprising the at least one object boundary.
2. The method of claim 1 , wherein the output is provided in real-time.
3. The method of claim 1 , wherein the deep neural network has sufficient depth to produce a saliency map for the input image with sufficient resolution for an application utilizing the output.
4. The method of claim 1 , wherein the deep neural network has a structure analogous to active contour energy constraints, to produce active contours for objects through a minimization of energy cost functions, using the deep neural network.
5. The method of claim 1 , wherein the output is used for any one of: object detection, localization, lane or roadway identification, or segmentation for computer vision applications.
6. The method of claim 1 , being applied in place of an existing active contour application to provide training and scalability from the deep neural network implementation.
7. The method of claim 1 , for enabling real-time active contour applications related to real-world images and/or videos.
8. The method of claim 1 , wherein the output is used to detect and/or segment vehicles and/or pedestrians for traffic analysis.
9. The method of claim 1 , wherein the output is used to detect and/or segment normal and/or abnormal medical structures from imagery or video, in both visible and non-visible bands.
10. The method of claim 1 , wherein the output is used to provide industrial quality control by detecting defects or verifying components in a manufacturing process.
11. The method of claim 1 , wherein the output is used to identify and classify other vehicles in an autonomous vehicle.
12. The method of claim 1 , wherein the output is used to track, identify, classify and/or count pedestrians.
13. The method of claim 1 , wherein the output is used to detect or track objects moving through a warehouse.
14. The method of claim 1 , wherein the output is used to identify behaviour in a surveillance system.
15. A non-transitory computer readable medium comprising computer executable instructions for detecting object boundaries, comprising instructions for:
obtaining an input image;
processing the input image using a series of input layers in a deep neural network to obtain a set of global features;
using outputs of a plurality of the input layers to perform edge detection at different scales and combining the outputs to obtain a set of local features;
combining the global and local features and applying an optimization function that has been trained to minimize intra-class variance and maximize inter-class variance according to one or more active contour energy constraints, to enhance at least one object boundary from the input image; and
providing an output comprising the at least one object boundary.
16. A computer vision system comprising:
an engine configured for receiving images and/or video; and
a memory comprising computer executable instructions for detecting object boundaries, comprising instructions for:
obtaining an input image;
processing the input image using a series of input layers in a deep neural network to obtain a set of global features;
using outputs of a plurality of the input layers to perform edge detection at different scales and combining the outputs to obtain a set of local features;
combining the global and local features and applying an optimization function that has been trained to minimize intra-class variance and maximize inter-class variance according to one or more active contour energy constraints, to enhance at least one object boundary from the input image; and
providing an output comprising the at least one object boundary.
17. The system of claim 16 , wherein the deep neural network has sufficient depth to produce a saliency map for the input image with sufficient resolution for an application utilizing the output.
18. The system of claim 16 , wherein the deep neural network has a structure analogous to active contour energy constraints, to produce active contours for objects through a minimization of energy cost functions, using the deep neural network.
19. The system of claim 16 , being applied in place of an existing active contour application to provide training and scalability from the deep neural network implementation.
20. The system of claim 16 , for enabling real-time active contour applications related to real-world images and/or videos.