By capturing multiple images, either sequential images from a single camera sensor, or from multiple sensors concurrently, valuable extra information is obtained. Using the extra information and extra computational processing power, we can generate images not previously possible. In some cases this may be new levels of quality, or it can also be new camera features.
Deep learning applied in the camera signal processing pipeline can provide new levels of performance for certain problems, but there are multiple challenges. Ground truth training data is difficult to capture or compute, expensive to assemble, and sometimes ineffective. For intermediate points in the camera pipeline, ground truth may be impossible to capture. Additionally, training data may change from image sensor to sensor.