Operator Quantization for Low-Power Image Segmentation Deployment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation models require high hardware costs and computational power, making it difficult to deploy them on low-cost hardware devices and hindering the widespread application of high-precision image recognition technologies.
Innovation Solution
A method for quantizing image segmentation models by evaluating computational power, selecting and optimizing operators based on a threshold, and converting the model into a target model suitable for deployment on low-cost hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-precision image segmentation models are deployed, then segmentation accuracy is improved, but hardware cost and computational power requirements increase
Solution Approach 1:
The patent applies quantization technology to change the precision parameters of model operators from high precision (e.g., FP32) to low precision (e.g., INT8), thereby reducing computational power requirements and hardware costs while maintaining acceptable segmentation accuracy. This parameter transformation enables deployment on low-cost hardware devices.
2Measurement precision
If high-precision image segmentation models are deployed, then segmentation accuracy is improved, but device cost increases
Solution Approach 1:
By changing the precision parameters of model operators through quantization, the patent reduces the hardware specifications required for deployment. This parameter transformation allows high-precision segmentation models to run on low-cost hardware devices, thereby reducing device manufacturing costs while maintaining segmentation accuracy.
3Measurement precision
If segmentation algorithms are deployed on embedded systems, then image segmentation capability is achieved, but processing time increases
Solution Approach 1:
The patent quantizes model operators to low precision, which significantly reduces the computational complexity and processing time of image segmentation operations on embedded systems. This parameter change enables real-time or near-real-time segmentation while maintaining acceptable accuracy, thereby reducing the time loss associated with deployment on resource-constrained devices.
Data Source
AI summary
A computational power evaluation result of a device is obtained. The device is to be deployed with an image segmentation model, the computational power evaluation result indicates operational performance of the device. The image segmentation model includes a plurality of operators. At least a first operator is selected from the plurality of operators based on the computational power evaluation result, a data processing duration of the first operator on the device exceeds a threshold. A quantization processing is performed on the first operator based on a difference between the data processing duration of the first operator and a desired processing duration, to obtain a first quantization operator, a data processing duration of the first quantization operator on the device is less than the desired processing duration. Based on at least the first quantization operator, the image segmentation model is converted into a target model for deployment onto the device.