Image Inference Thinning Filters for Low-Power High-Resolution AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI processing systems face challenges in achieving high-resolution image processing while maintaining low power consumption, particularly in edge computing applications, where conventional methods like pruning and quantization of AI models are insufficient for further power efficiency improvements.
Innovation Solution
An inference device that employs thinning processing to reduce the amount of convolution arithmetic operations by thinning image data in a spatial direction without overlapping pixels, using a convolution arithmetic operation unit to apply filters to thinned data and a time-series filter for inference, thereby reducing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional AI processing is performed on high-definition images, then processing performance is improved, but power consumption increases
Solution Approach 1:
The patent extracts and removes redundant information from image data through thinning processing. By eliminating duplicate or less important pixel information while preserving essential features, the system reduces the data volume requiring AI processing, thereby lowering computational load and power consumption while maintaining processing performance.
Solution Approach 2:
The patent changes the parameter of image data density by applying thinning processing. This transforms full-resolution image data into thinned image data with reduced pixel density, directly reducing the amount of convolution arithmetic operations needed and consequently lowering power consumption while preserving AI processing capability.
2Measurement precision
If image resolution is increased to improve processing accuracy, then measurement precision is improved, but the amount of convolution arithmetic operation increases
Solution Approach 1:
The patent extracts only the essential information needed for accurate AI processing by removing redundant pixels through thinning processing. This maintains the necessary measurement precision for AI tasks while significantly reducing the complexity of convolution operations by decreasing the number of pixels that require processing.
3Use of energy by moving object
If pruning or quantization is applied to reduce AI model weight, then power consumption is reduced, but required performance is difficult to realize
Solution Approach 1:
The patent performs thinning processing as a preliminary action before AI inference. By pre-processing the image data to remove redundant information, the system reduces the input data volume for the AI model, which decreases computational requirements and power consumption without requiring modifications to the AI model itself, thus maintaining performance.
Data Source
AI summary
An inference device including an input unit that inputs image data continuous in a predetermined direction; a thinning processing unit that executes thinning of a plurality of pieces of image data input by the input unit such that a pixel array indicating a plurality of patterns is repeated and pixels in a spatial direction do not overlap, and outputs a plurality of pieces of thinned image data; a convolution arithmetic operation unit that performs a convolution arithmetic operation by applying a plurality of thinning filters configuring the plurality of patterns divided from a filter having a weighting factor to each of the plurality of pieces of thinned image data; an inference unit that executes inference regarding the plurality of pieces of image data using a time-series filter on a basis of a plurality of convolution arithmetic operation results; and an output unit that outputs an inference result.