Techniques for context and performance adaptive processing in ultra low-power computer vision systems

摘要

Techniques are disclosed to adaptively control an image sensor and image signal processor (ISP), collectively known as an image pipeline, such that a minimum amount of power is consumed by an electronic device while producing images that can be accurately analyzed by a computer vision application. The techniques disclosed herein can be implemented in various electronic devices capable of capturing and processing image data, such as, for example, smart phones, wearable computers, laptops, tablets, and other mobile computing or imaging systems. In an embodiment, the techniques disclosed herein are implemented in a system-on-chip device.

主权项

1. An image processing system comprising:
an image sensor; an image signal processor communicatively coupled to the image sensor; a memory storing a configuration parameter of at least one of the image sensor and the image signal processor, the configuration parameter controlling performance and power consumption of the at least one of the image sensor and the image signal processor; a signal metrics processor communicatively coupled to the image sensor and the image signal processor, the signal metrics processor configured to measure a characteristic of captured image data; and a computer vision controller communicatively coupled to the signal metrics processor and configured to initiate capturing of image data in accordance with a selected computer vision task, wherein the computer vision controller is configured to receive information descriptive of performance requirements of the selected computer vision task and to adjust the configuration parameter of the at least one of the image signal processor and the image sensor based in part on the information descriptive of the performance requirements of the selected computer vision task, thereby causing the at least one of the image sensor and the image signal processor to consume power varying with the performance requirements of the selected computer vision task.