Autonomous Driving Image Signal Processor Selection via Perception Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image signal processors in autonomous driving systems adjust image quality based on human preferences, which does not effectively enhance the recognition accuracy and safety of the system.
Innovation Solution
A method and device for constructing an image signal processor by creating multiple candidate processors using various image signal processing modules, processing original images, and selecting the best one based on preset performance criteria through perception task models to meet the system's requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If image quality adjustment is based on human eye preferences, then image quality meets human preferences, but recognition accuracy of perception module deteriorates
Solution Approach 1:
Instead of optimizing image processing for human visual preferences, the patent inverts the approach by optimizing for machine perception requirements. The ISP construction uses perception task models to evaluate and select processing parameters that maximize recognition accuracy for autonomous driving scenarios, rather than prioritizing human aesthetic preferences.
Solution Approach 2:
The patent implements a feedback mechanism where perception task models evaluate the output images from candidate ISPs and provide performance metrics. This feedback loop allows the system to iteratively optimize ISP parameters based on actual perception task performance, ensuring that image processing improvements directly translate to better recognition accuracy.
2Measurement precision
If multiple candidate ISPs are constructed and evaluated, then recognition accuracy improves, but device complexity increases
Solution Approach 1:
The patent segments the ISP construction process into distinct modules: candidate ISP generation, image processing, perception task evaluation, and selection. This modular segmentation allows each component to be independently optimized and managed, reducing overall system complexity despite evaluating multiple candidates.
Solution Approach 2:
The patent performs preliminary action by pre-construction of multiple candidate ISPs and their evaluation using perception task models. This offline preparation allows the system to select the optimal ISP configuration in advance, reducing runtime complexity while maintaining high recognition accuracy through pre-validated configurations.
Data Source
Figure 1~2
AI summary
The disclosure provides a method for constructing an image signal processor, a device, an autonomous driving system, and a medium. The method includes: constructing a plurality of candidate image signal processors based on at least some of the image signal processing modules from a collection; each candidate image signal processor processing an acquired original image, to generate a plurality of sets of processed target images; inputting each set of target images into at least one preset perception task model, and outputting a first perception result corresponding to each perception task model; and selecting the candidate image signal processor that meets a preset performance criteria as the target image signal processor based on all the first perception results corresponding to each candidate image signal processor, thereby realizing the selection of the target image signal processor from the perspective of perception of an autonomous driving system, reducing the error rate in evaluation acquired based on human eyes, improving the recognition accuracy of the autonomous driving system, and further ensuring safe operation of the autonomous driving system.