Low-light image selection for neural network training

A machine learning system comprising:

• a processing system;
• and
• a memory storing instructions that, when executed by the processing system, causes the machine learning system to:
• obtain a series of images of a scene, each of the images including a plurality of pixel values;
• determine a number of pixel values of the plurality of pixel values in each image that are equal to a pixel saturation value;
• select a representative image from the series of images based at least in part on the number of pixel values in each image that are equal to the pixel saturation value;
• and
• train a neural network to reproduce the representative image based on a plurality of remaining images in the series, the plurality of remaining images in the series not including the representative image.

A method of training neural networks, comprising:

• obtaining a series of images of a scene, each of the images including a plurality of pixel values;
• determining a number of pixel values of the plurality of pixel values in each image that are equal to a pixel saturation value;
• selecting first and second representative images from the series of images based at least in part on the number of pixel values in each image that are equal to the pixel saturation value;
• generating a ground truth image based on the selected representative images;
• and
• training a neural network to reproduce the ground truth image based on a plurality of the images in the series.