Image Dimension Reduction via Channel Filtering for AI Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Mid-tier/low-end computing devices face challenges in performing AI-based use cases with high-resolution images due to high memory and computational requirements, leading to accuracy degradation when downscaling images or simplifying neural networks, which results in either prolonged inference times or poor accuracy.
Innovation Solution
Transforming images from a spatial domain to a non-spatial domain, filtering out irrelevant channels using AI engines, and embedding a generic stub layer to bypass unnecessary neural network layers, allowing for dimensionality reduction without retraining the network, thus enabling faster and more accurate AI-based use cases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution images are used for AI-based use cases, then accuracy is improved, but computational requirements and memory usage increase
Solution Approach 1:
The image processing pipeline is segmented into multiple stages: initial downscaling, selective channel processing, and targeted neural network layer application. This allows the system to process images at reduced resolution for most operations while maintaining high-resolution processing only where necessary for accuracy-critical tasks.
Solution Approach 2:
Different processing qualities are applied to different parts of the image data. The patent applies full resolution processing only to specific channels (e.g., luminance channel) that contribute most to accuracy, while using reduced resolution for other channels (e.g., chrominance channels), thereby optimizing the balance between accuracy and computational load.
2Productivity
If image resolution is downsampled to reduce computation, then inference time is reduced, but accuracy degrades
Solution Approach 1:
The patent applies different resolution qualities to different channels based on their importance. The luminance channel is processed at full resolution to preserve edge and detail information critical for accuracy, while chrominance channels are processed at lower resolution since they contribute less to overall accuracy but still provide useful information.
Solution Approach 2:
Instead of processing all channels at full resolution, the patent applies partial processing - using full resolution only where necessary (luminance channel) and reduced resolution elsewhere (chrominance channels). This partial action approach maintains sufficient accuracy while significantly reducing computational load.
3Adaptability or versatility
If neural network complexity is reduced to enable deployment on mid-tier devices, then device compatibility is improved, but use case performance degrades
Solution Approach 1:
The neural network is segmented into different processing stages with varying complexity. Early layers use simplified operations on downsampled inputs, while later layers apply more complex operations on selectively processed channels. This segmentation allows the network to be adapted to mid-tier devices while maintaining performance through strategic use of computational resources.
Solution Approach 2:
Different levels of neural network processing are applied to different channels. The patent uses full-resolution processing for the luminance channel through the network and reduced-resolution processing for chrominance channels, ensuring that accuracy-critical pathways maintain high quality while other pathways provide supplementary information with lower computational cost.
Data Source
AI summary
Embodiments of the disclosure provide a method and device for efficiently reducing dimensions of an image frame by an electronic device. The method includes: receiving the image frame; transforming the image frame from a spatial domain comprising a first plurality of channels to a non-spatial domain comprising a second plurality of channels, where a number of the second plurality of channels is greater than a number of the first plurality of channels; removing channels comprising irrelevant information from among the second plurality of channels using an AI engine to generate a low-resolution image frame in the non-spatial domain; and providing the low-resolution image frame to a neural network for a faster and accurate inference of the image frame.