Neural network image processing

1. A computer, comprising:

a processor; and

a memory, the memory including instructions executable by the processor to:

determine a second convolutional neural network (CNN) training dataset by:

determining an underrepresented object configuration and an underrepresented noise factor corresponding to an object in a first CNN training dataset;

generating one or more simulated images including the object corresponding to the underrepresented object configuration in the first CNN training dataset by inputting ground truth data corresponding to the object into a photorealistic rendering engine; and

generating one or more synthetic images including the object corresponding to the underrepresented noise factor in the first CNN training dataset by processing the simulated images with a generative adversarial network (GAN);

train the CNN using the first and the second CNN training datasets; and

input an image acquired by a sensor to the trained CNN and output an object label and an object location corresponding to the object corresponding to the underrepresented object configuration and underrepresented object noise factor; wherein the trained CNN is executed on a second computer in a vehicle and the object label and the object location are used by the second computer to operate the vehicle by controlling one or more of vehicle propulsion, vehicle brakes or vehicle steering.

2. The computer of claim 1 , the instructions including further instructions to:

determine a second underrepresented noise factor that corresponds to the CNN not outputting the object label and the object location corresponding to the object by testing the CNN with a test dataset based on real world images;

generate one or more synthetic images including the object corresponding to the second underrepresented noise factor to determine a third CNN training dataset by inputting images from the first training dataset and the second training dataset into the GAN; and

retrain the CNN using the third training dataset to output a second object label and a second object location corresponding to the second underrepresented noise factor using the first, second and third CNN training datasets.

3. The computer of claim 1 , wherein the first object configuration is underrepresented when a number of images in the first CNN training dataset that include the first object configuration is less than an average number of images that include each other object configuration.

4. The computer of claim 1 , wherein the first noise factor is underrepresented when a number of images in the first CNN training dataset that include the first noise factor is less than an average number of images that include each other noise factor.

5. The computer of claim 1 , wherein the first and second CNN training datasets include images that include the object and corresponding ground truth that includes the object label and the object location for the object included in the images.

6. The computer of claim 1 , wherein training a second CNN with the second CNN training dataset reduces false positives output by a first CNN trained with the first CNN training dataset when processing real-world data, and wherein a false positive is an object label incorrectly applied to an object occurring in an image.

7. The computer of claim 1 , wherein the object label is a text string that identifies an object included in an input image and the object location is a bounding box corresponding to the object included in the input image.

8. The computer of claim 1 , wherein object configurations include values and an arrangement of pixels corresponding to the object based on an object location, an object orientation, and partial obscuring of the object by another object and/or another object's shadow.

9. The computer of claim 1 , wherein noise factors include values and an arrangement of pixels corresponding to the object based on environmental conditions including one or more of partial sunlight, full sunlight, rain, snow, fog, and dust.

10. The computer of claim 9 , wherein the photorealistic rendering engine is a software program that inputs ground truth data regarding object configuration and outputs an image.

11. The computer of claim 1 , wherein the GAN is trained to generate simulated images by training the GAN with real-world images that include noise factors.

12. The computer of claim 11 , wherein training the GAN with real-world images includes comparing an image output by the GAN with input real-world images to determine similarity between the image output by the GAN with input real-world images by correlating the output images with the input real-world images.

13. The computer of claim 1 , wherein the CNN includes convolutional layers that output hidden variables to fully-connected layers output states that include the object label and object location.

14. The computer of claim 1 , wherein the CNN is trained to output states corresponding to the object label and object location by processing an input image a plurality of times and comparing the output states to ground truth data corresponding to the input image.

15. The computer of claim 1 , wherein the vehicle is operated by the second computer by controlling one or more of a powertrain controller, a brake controller, and a steering controller.

16. The computer of claim 15 , wherein operating the vehicle includes determining a vehicle path by a second computer that avoids contact with the object.

17. A method, comprising:

determining a second convolutional neural network (CNN) training dataset by:

determining an underrepresented object configuration and an underrepresented noise factor corresponding to an object in a first CNN training dataset;

generating one or more simulated images including the object corresponding to the underrepresented object configuration in the first CNN training dataset by inputting ground truth data corresponding to the object into a photorealistic rendering engine; and

generating one or more synthetic images including the object corresponding to the underrepresented noise factor in the first CNN training dataset by processing the simulated images with a generative adversarial network (GAN) to determine a second CNN training dataset;

training the CNN using the first and the second CNN training datasets; and

inputting an image acquired by a sensor to the trained CNN and outputting an object label and an object location corresponding to the underrepresented object configuration and underrepresented object noise factor wherein the object label and the object location are used to operate a vehicle by controlling one or more of vehicle propulsion, vehicle brakes or vehicle steering.

18. The method of claim 17 , further comprising:

determining a second underrepresented noise factor that corresponds to the CNN not outputting the object label and the object location corresponding to the object by testing the CNN with a test dataset based on real world images;

generating one or more synthetic images including the object corresponding to the second underrepresented noise factor to determine a third CNN training dataset by inputting images from the first CNN training dataset and the second CNN training dataset into the GAN; and

retraining the CNN using the third CNN training dataset to output a second object label and a second object location corresponding to the second underrepresented object noise factor using the first, second, and third CNN training datasets.

19. The method of claim 17 , wherein a first object configuration is underrepresented when a number of images in the first CNN training dataset that include the first object configuration is less than an average number of images that include each other object configuration.

20. The method of claim 17 , wherein a first noise factor is underrepresented when a number of images in the first CNN training dataset that include the first noise factor is less than an average number of images that include each other noise factor.