Network adaptive semi-precision quantification image processing method and system

Abstract

The invention provides a network adaptive semi-precision quantification image processing method and system. The method comprises the following steps: collecting image data of edge computing equipment;inputting the image data into a pre-established depth residual convolution quantization network for image processing for processing to obtain the category of a target corresponding to the image data,positioning and the category of pixels in the image; wherein the deep residual convolutional quantization network is obtained by training based on a deep network adaptive semi-precision quantizationmethod, and the semi-precision quantization is carried out by adopting a floating-point number with a half digit. Semi-precision adaptive quantization is carried out on the deep residual convolution quantization network, the size of the deep residual convolution quantization network can be reduced on the premise that the precision is guaranteed, and the requirement for computing resources is lowered.

Description

Technical field [0001] The invention belongs to the technical field of digital image processing and computer vision, and specifically relates to an image processing method and system for network adaptive semi-precision quantization. Background technique [0002] With the development of deep learning, convolutional neural networks have been widely used in the field of target detection. In order to ensure the accuracy of target detection, target detection models based on convolutional neural networks are becoming more and more complex, which makes the parameter of the model increase exponentially, which not only increases the model storage space, but also increases the computational cost accordingly. This makes the current high-precision target detection models unable to be deployed on edge devices with limited resources. The quantization method can convert the 32-bit single-precision floating-point number operation used by the deep learning model into a low-bit depth numerical ty...

AI Technical Summary

Problems solved by technology

However, this method weakens the expressive ability of the network. Even for complex classification tasks, the performance will drop a lot, and it cannot be applied to more complex detection models.

The paper "TrainedTernary Quantization" is a typical three-value quantization method. This method does not simply quantize parameters to 0, +1, -1, but quantizes each layer into different parameters by learning, although certain The performance loss can be avoided to a certain extent, but this method only quantizes the weight value, not the activation value, and still uses 32-bit floating point numbers to represent the quantization parameters, which is also not suitable for deployment on edge devices

In addition, there are some post-training quantization algorithms that directly quantify model parameters for inference. Although this method is simple to implement, it does not retrain the network to learn and correct the errors caused by quantization, which leads to greater performance loss.

Images