Weight-Feature DNN Quantization for Resource-Limited Image Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural network models for image classification/detection require significant hardware resources and have high computational demands, which are not feasible for devices with limited computing capabilities and memory, and current quantization methods often compromise accuracy.
Innovation Solution
A method that quantizes deep neural network models based on weight features, determining a tailored quantization policy to minimize resource usage and time consumption while maintaining accuracy, by analyzing per-layer and per-channel quantization errors and clustering weight channels to optimize quantization units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural network models are used for image classification/detection, then processing accuracy is improved, but hardware resource consumption and computational demand increase
Solution Approach 1:
The patent applies parameter changes by transforming the data precision parameters of the deep neural network model from high-precision floating-point format to low-precision quantized format. This parameter transformation reduces the bit-width of weight and activation data, thereby decreasing memory usage and computational complexity while maintaining acceptable accuracy levels for image classification and detection tasks
2Volume of stationary object
If model size is reduced through quantization, then hardware demand decreases, but quantization process time increases and model performance may be affected
Solution Approach 1:
The patent applies preliminary action by performing quantization calibration in advance during the model preparation phase. Statistical information such as minimum and maximum values of activation data are collected and stored beforehand, which enables the quantized model to be executed efficiently without performing complex quantization calculations during real-time inference, thus reducing inference time
Solution Approach 2:
The patent applies partial action by selectively applying different quantization strategies to different parts of the network. Per-layer quantization is applied where appropriate to reduce overhead, while per-channel quantization is used where it provides significant accuracy benefits. This selective approach balances quantization time and model performance
3Measurement precision
If per-channel quantization is applied, then model accuracy is maintained, but quantization complexity and time consumption increase
Solution Approach 1:
The patent applies dynamics by implementing an adaptive quantization strategy that dynamically selects between per-layer and per-channel quantization based on the specific characteristics of each layer. The system evaluates whether per-channel quantization is necessary for each layer, allowing the quantization approach to adapt to the actual needs of different network layers rather than uniformly applying the most complex method throughout
Data Source
AI summary
An image processing method, apparatus and device, and a readable storage medium are disclosed, including: obtaining a target image; inputting the target image into a quantized target deep neural network model for classification/detection to obtain an output result; and processing the target image according to a policy corresponding to the output result. A process of performing quantization to obtain the target deep neural network model includes: obtaining a pre-trained floating point type deep neural network model; extracting weight features of a deep neural network model; determining a quantization policy using the weight features; and quantizing the deep neural network model according to the quantization policy to obtain the target deep neural network model.