A convolutional neural network accelerator circuit based on a fast filtering algorithm

Abstract

The invention discloses a convolutional neural network accelerator circuit based on a fast filtering algorithm. In order to reduce the calculation amount of a convolutional neural network algorithm (CNN), the method provided by the invention utilizes a fast filtering algorithm to eliminate the redundancy of overlapped region calculation between convolutional windows in two-dimensional convolutionoperation, so that the algorithm strength is reduced, and the convolution calculation efficiency is improved. Next, a convolution calculation acceleration unit of a four-parallel fast filtering algorithm is designed, and the unit is realized by adopting a parallel filtering structure which is composed of a plurality of small filters and is low in complexity. For the programmable FPGA design, not only can the consumption of hardware resources be reduced, but also the running speed can be increased. Meanwhile, the activation function is subjected to optimization design, and a hardware circuit ofthe activation function (sigmoid) is designed by using a piecewise fitting method combining a lookup table and a polynomial, so that the hardware circuit of the approximate activation function is ensured not to reduce the precision.

Description

technical field [0001] The invention belongs to the field of digital signal processing and digital integrated circuit design, and more specifically relates to a convolutional neural network accelerator circuit based on a fast filtering algorithm. Its main application fields include image recognition, image filtering, and image compression. Background technique [0002] Convolutional neural network (CNN) is a deep learning algorithm originated from artificial neural network, which has a strong ability to resist displacement and deformation interference in image processing. At the same time, due to the remarkable achievements of CNN in solving advanced abstract cognitive problems, it has been more and more widely used in image classification, pattern recognition, face detection and other fields. Among them, the recognition of handwritten characters in images often uses convolutional neural network algorithms. In the convolutional neural network algorithm, the convolution oper...

AI Technical Summary

Problems solved by technology

For example, on the basis of the direct convolution structure, some literature proposes to add enough computing resources to perform convolution calculation in parallel to generate a row of output feature maps. Although this method improves the calculation performance (GOPS), this calculation mode reduces the utilization of computing units while consuming a lot of hardware resources

Images