Dynamic Vision-Guided Image Compression for Faster Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks in safety-critical systems like autonomous vehicles require accurate and rapid classification probabilities, which are challenging due to the need for real-time decision-making and handling of moving obstacles.
Innovation Solution
A method involving a Dynamic Vision Sensor and an image sensor capturing overlapping views, applying higher image compression to non-movement areas while maintaining minimal compression on movement areas to enhance classification accuracy and speed using a machine learning algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If higher image compression is applied to the entire image, then processing time is reduced and data transmission is optimized, but classification accuracy deteriorates due to loss of detail in movement areas
Solution Approach 1:
The patent applies different compression levels to different regions of the image based on movement detection. Static regions receive high compression to reduce data size and processing time, while dynamic regions maintain low compression to preserve detail for accurate classification. This local differentiation resolves the contradiction by optimizing each region according to its specific requirements.
Solution Approach 2:
The image is segmented into movement areas and non-movement areas based on analysis from the dynamic vision sensor. This segmentation allows the system to apply appropriate compression strategies to each segment independently, ensuring that classification accuracy is maintained in critical movement regions while optimizing processing efficiency in static regions.
2Measurement precision
If no image compression is applied, then classification accuracy is maintained, but processing time increases and real-time decision-making capability deteriorates
Solution Approach 1:
Instead of applying uniform compression or no compression to the entire image, the system applies local quality differentiation where static regions are highly compressed for speed optimization while dynamic regions remain uncompressed for accuracy. This resolves the contradiction by recognizing that different parts of the image have different temporal and accuracy requirements.
Solution Approach 2:
The compression strategy is dynamically adjusted based on real-time movement detection. The system continuously identifies movement regions and adjusts compression levels accordingly, creating a dynamic processing pipeline that adapts to changing scene conditions to balance accuracy and speed requirements.
3Measurement precision
If the entire image is processed at high resolution, then classification accuracy is improved, but computational resources are wasted on static areas that do not require detailed analysis
Solution Approach 1:
The system applies local quality differentiation where computational resources are concentrated only on dynamic regions that require detailed analysis for classification. Static regions are processed at lower resolution, significantly reducing the total computational load while maintaining accuracy where it matters most.
Solution Approach 2:
The system extracts and prioritizes processing resources toward movement regions identified by the dynamic vision sensor. By taking out the assumption that the entire image requires uniform high-resolution processing, the system efficiently allocates computational resources to only those areas containing relevant information for classification decisions.
Data Source
AI summary
A method comprising the steps of: receiving an image from a dynamic vision sensor and an image from an image sensor, wherein the dynamic vision sensor and the image sensor capture at least an overlapping field of view; determining, from the image received from the dynamic vision sensor, an area of movement in the field of view of the image from the image sensor; applying a higher level of image compression to the areas of no movement compared with the area of movement in the image from the image sensor to produce a processed image; and outputting the processed image to a neural network.