Visibility-Adaptive Neural Image Analysis for Edge Cameras
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
ANN-based image analysis solutions are computationally demanding and often unsuitable for edge devices due to limited resources, especially when multiple tasks are required under varying visibility conditions such as fog and smog.
Innovation Solution
Adaptive selection of ANN architectures based on visibility conditions, using lower-resolution architectures for reduced visibility scenarios to conserve computational resources, and optionally downscaling images to match the selected architecture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a higher-resolution ANN architecture is used for image analysis, then analysis accuracy is improved, but computational complexity and resource consumption increase
Solution Approach 1:
The system dynamically selects between different ANN architectures (higher-resolution and lower-resolution) based on real-time visibility conditions. When visibility is good, the higher-resolution architecture is used for maximum accuracy. When visibility deteriorates (fog, smog), the system switches to the lower-resolution architecture to reduce computational complexity while maintaining adequate performance.
Solution Approach 2:
The invention changes the resolution parameter of the ANN architecture based on visibility conditions. By adjusting the input resolution parameter to match the actual scene visibility, the system optimizes the balance between analysis accuracy and computational resource consumption, avoiding unnecessary processing when details are obscured by poor visibility.
2Reliability
If a higher-resolution ANN architecture is used, then detection capability is improved, but resource consumption increases
Solution Approach 1:
The system dynamically adapts its computational resources to match the actual detection needs imposed by visibility conditions. In clear conditions, full detection capability is utilized with higher-resolution processing. In poor visibility, the system reduces resolution to match the reduced information content in the scene, thereby conserving energy and computational resources.
Solution Approach 2:
The invention applies partial action by using only the necessary resolution level required for effective detection under current visibility conditions. Instead of always using maximum resolution, the system applies just enough processing power to achieve reliable detection, avoiding excessive resource consumption when high detail is not available in the input image.
3Adaptability or versatility
If multiple ANN architectures are maintained for different visibility conditions, then adaptability is improved, but device complexity increases
Solution Approach 1:
The system segments the ANN architectures into distinct versions (higher-resolution and lower-resolution) optimized for different visibility conditions. This segmentation allows the edge device to select the appropriate architecture based on current conditions, achieving adaptability while keeping each individual architecture relatively simple and manageable.
Solution Approach 2:
The invention creates a universal system that can handle multiple visibility conditions using a set of ANN architectures. Each architecture serves multiple purposes: the higher-resolution architecture handles clear conditions and can downscale for moderate conditions, while the lower-resolution architecture handles poor visibility conditions, providing multi-functionality across different environmental scenarios.
Data Source
Figure 1
Figure 2~3
Figure 4A~5
AI summary
A solution (200) for analyzing images of a scene captured under different visibility conditions is provided. The solution includes obtaining images (212) of a scene captured by one or more cameras (210) and, for each image, obtaining (220) an indication of an actual or assumed visibility (distance) at the scene when the image was captured; selecting (230), based on the visibility, an artificial neural network (ANN) architecture from a plurality (240) of ANN architectures, wherein the ANN architectures are each trained for image analysis but configured for different input image resolutions, and analyzing the image using the selected ANN architecture. If the selected ANN architecture has a lower input image resolution than the image, the image may be downscaled to match the input image resolution of the selected ANN architecture. There is provided a corresponding method, device (e.g. camera), computer program and computer program product.