AI Processor Neural Network Downsampling Upsampling Power
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies using artificial neural networks on terminal devices face challenges with high resource and power consumption, especially when handling large images or requiring real-time processing, which is exacerbated by the need for efficient on-device processing.
Innovation Solution
An AI processor that adaptively operates by downsampling and upsampling images using a neural network model trained for both processes, reducing resource and power consumption through efficient encoding and decoding methods, and dynamically adjusting neural network operations based on image size and processing capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If image processing is performed using a neural network on terminal devices, then image quality improvement and processing capability are enhanced, but resource and power consumption increase significantly
Solution Approach 1:
The image processing task is segmented into multiple stages: downsampling stage, encoding stage, and upsampling stage. Each stage is processed by dedicated neural network operators that can be independently controlled and optimized, allowing the system to process only necessary portions of the image at full resolution while maintaining overall processing capability.
Solution Approach 2:
The system dynamically adjusts processing parameters including clock frequencies of neural network operators based on input image characteristics and processing requirements. This dynamic adaptation allows the processor to optimize power consumption by allocating computational resources only when and where needed, rather than operating at full capacity continuously.
2Manufacturing precision
If the size of image increases, then processing detail and quality improve, but computation time and resource consumption increase
Solution Approach 1:
The image is divided into different resolution zones: full-resolution regions requiring detailed processing and downsampled regions for broader context. This segmentation allows the system to apply computationally intensive operations only where high detail is necessary, reducing overall computation time while preserving critical image quality.
Solution Approach 2:
Different processing quality levels are applied to different regions of the image based on local requirements. Critical regions receive full-resolution neural network processing for high detail, while less critical regions use downsampled processing, optimizing the balance between overall image quality and computation time.
3Speed
If real-time processing is required, then responsiveness improves, but power consumption and computational load increase
Solution Approach 1:
The system performs downsampling and preprocessing operations before the main neural network processing to reduce the computational burden of real-time operations. By preparing data in advance at lower resolution, the system can achieve real-time responsiveness with reduced power consumption during critical processing phases.
Solution Approach 2:
The neural network operators operate in periodic cycles with adjustable clock frequencies, allowing the system to burst-process when needed for real-time responsiveness while entering lower-power states between operations. This periodic operation pattern maintains responsiveness while reducing average power consumption.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An artificial intelligence (Al) processor includes at least one memory; a plurality of neural network operators comprising circuitry configured to process an image; and a controller configured to control the at least one memory and the plurality of neural network operators. The controller controls input image data of an image to be stored in the at least one memory and controls at least one of the plurality of neural network operators to perform a neural network operation on image data split based on a size of the image and data processing capabilities of the plurality of neural network operators, and output upscaled image data.