Quantitative neural network acceleration method based on field programmable array

Abstract

The invention discloses a quantitative neural network acceleration method based on a field programmable array, is applied to the field of image processing, and aims to solve the problem of low image processing efficiency in the prior art. Each layer of a neural network for image processing is expressed as a calculation graph, after input and weight are subjected to convolution or full-connection calculation, a bias value is added, and a final output is obtained through an activation function; the weight space is approximated to a sparse discrete space; numerical quantification is performed on the processed weight to obtain a quantified neural network for image processing; then an accelerator matched with the quantitative neural network for image processing is designed; and each layer of the quantized image processing neural network is calculated according to the corresponding accelerator to obtain an image processing result. By adopting the method provided by the invention, the image processing application can be deployed in a resource-limited embedded system, and the method has the characteristics of rapid reasoning and low power consumption.

Description

technical field [0001] The invention belongs to the field of image processing, in particular to an image processing technology based on a neural network. Background technique [0002] Neural networks (NN) have achieved good results in many fields such as object detection and semantic segmentation, but how to deploy artificial intelligence (AI) applications of neural networks in practical applications such as autonomous driving and autonomous robots is challenging. This is because the devices in practical applications are generally resource-constrained embedded systems, which have less memory and insufficient computing power, but neural networks usually have a huge amount of parameters and calculations, which require a large amount of storage resources and Computing resources, resource-constrained embedded systems cannot be satisfied. The low-precision quantization of the neural network model can effectively reduce storage requirements. Using field programmable gate arrays a...

AI Technical Summary

Problems solved by technology

Generally speaking, the CPU can complete 10-100 GFLOP operations per second, but the energy efficiency is usually lower than 1GOP / J, so it is difficult to meet the high performance requirements of the neural network and the low energy consumption requirements of the mobile terminal.

In contrast, the peak performance provided by the GPU can reach 10TOP / S, so it is an excellent choice for high-performance neural network applications, but it cannot meet the low energy consumption requirements of the mobile terminal.

Images