Training a deep learning system for maritime applications

1. A method comprising:

training, using a processor, an object detection network with training images to identify and classify objects in images from a sensor system disposed on a maritime vessel, wherein the object detection network is a convolution neural network that includes layers having weights;

identifying objects in the images using the object detection network in an offline mode;

classifying the objects in the images using the object detection network in the offline mode such that pixels in the images are associated with at least one the objects using a filter, and wherein the objects include at least a body of water and a watercraft;

generating heat maps in the offline mode, wherein the heat maps are based on stereoscopic output disparity maps that depict distance with shading; and

sending instructions regarding operation of the maritime vessel, using the processor, based on the objects that are identified, wherein the instructions include a speed or a heading.

2. The method of claim 1 , wherein the method further comprising training, using the processor, the object detection network to send the instructions regarding operation of the maritime vessel based on the objects that are identified.

3. The method of claim 1 , wherein the objects further include a shore, an iceberg, a static far object, or a moving far object, and wherein the watercraft includes a personal non-powered vessel, recreational powered vessel, sailing yacht, cargo ship, cruise ship, coast guard boat, naval vessel, barge, tugboat, fishing vessel, workboat, under-powered vessel, or anchored vessel.

4. The method of claim 1 , further comprising performing deduplication of the training images using hash outputs prior to the training.

5. The method of claim 1 , wherein the training images are determined by:

receiving an initial image of a maritime object at a processor; and

changing at least one of water conditions, sky conditions, or sunlight in the image of the maritime object to generate one of the training images that includes the maritime object using the processor.

6. A non-transitory computer readable medium storing a program configured to instruct the processor to execute the identifying, the classifying, the generating, and the sending of claim 1 .

7. The object detection network trained using the method of claim 1 .

8. A method comprising:

receiving an image of a maritime object at a processor, wherein the maritime object is a watercraft;

changing at least one of water conditions, sky conditions, or sunlight in the image of the maritime object to generate a second image that includes the maritime object using the processor, wherein the changing includes:

generating a mask for the image of the maritime object;

synthesizing a background; and

blending the background and the image of the maritime object to form the second image; and

training an object detection network using the second image, wherein the object detection network is a convolution neural network that includes layers having weights.

9. The method of claim 8 , wherein a plurality of the second images are generated, wherein each of the plurality of the second images has a different one of the water conditions, the sky conditions, or the sunlight, and further comprising performing deduplication of the second images using hash outputs prior to the training.

10. The method of claim 8 , further comprising receiving a background image with the water conditions, the sky conditions, or the sunlight used in the second image.

11. The method of claim 8 , wherein the changing further includes blending a region of the image of the maritime object with a region of a background image.

12. The method of claim 8 , wherein the changing further includes determining a gradient field of the image of the maritime object.

13. The method of claim 12 , wherein the changing further includes determining a gradient field of a background image.

14. The method of claim 13 , wherein the changing further includes determining a gradient field of the second image fusing the image of the maritime object and the background image.

15. The method of claim 13 , wherein the changing further includes determining divergence of the second image.

16. The method of claim 15 , wherein the changing further includes determining a solution for a coefficient matrix.

17. The method of claim 8 , wherein the processor is in electronic communication with a data server, and wherein the data server provides the image of the maritime object.

18. A non-transitory computer readable medium storing a program configured to instruct a processor to execute the method of claim 8 .