Standard meal identification method and system based on INT4 quantization

Abstract

The invention discloses a standard meal identification method and system based on INT4 quantization, and belongs to the field of neural network models. The method comprises the following specific steps: S1, collecting and labeling standard meal image data, and constructing a training and calibration data set; s2, training an image classification model by using a floating-point number to obtain a floating-point number model to be quantized; s3, preprocessing the image classification model by using cross-layer equalization; quantizing the floating-point number model into an INT4 type model; s4, deploying an INT4 quantitative model to an MCU (Microprogrammed Control Unit); s5, integrating MCU camera data as quantitative model input; according to the method, the pipeline quantification neural network model is used, the precision loss of quantification of the neural network model can be effectively reduced, the simple standard meal identification requirement can be met, and the practicability is high; in addition, due to the fact that INT4 quantification has a large model compression proportion, the requirements for hardware memory and computing resources are reduced, and the method can be applied to a single-chip microcomputer with low cost and has high economic benefits.

Description

technical field [0001] The invention discloses a standard meal identification method and system based on INT4 quantification, and relates to the technical field of neural network models. Background technique [0002] In recent years, neural network models have been widely used in many fields and achieved very good results. However, due to the high model complexity and large size of the neural network model, the inference efficiency is low and the inference time is long. Therefore, how to design a model with low resource consumption, which can predict in real time and at the same time ensure the prediction accuracy has become a practical problem, especially in the application of simple scenarios. Standard meal recognition is a simple scene image recognition application, because the meals are relatively standard and there are relatively few types of meals, the image recognition accuracy rate is high, and it has high application value. However, due to the high complexity of t...

AI Technical Summary

Problems solved by technology

However, due to the high complexity of commonly used deep learning models for image classification, the hardware requirements are high, resulting in high hardware costs and power consumption.

Images