SNN-based image recognition software and hardware system

Abstract

The invention discloses an image recognition software and hardware system based on SNN, and the system consists of an input conversion unit, an intermediate calculation unit and an output recognition unit which are respectively used for completing the pulse sequence conversion of an input image, the information transmission of an intermediate pulse, and the final classification output of a model. The input conversion unit mainly comprises a preprocessing module and a filtering coding module, the intermediate calculation unit comprises a pulse transmission module, a competition suppression module and a weight learning module, and the output recognition unit comprises a global pooling module and a voting arbitration module. According to the invention, an end-to-end image recognition system solution based on a novel SNN self-classification model is realized, system distributed deployment and parallelism optimization about a heterogeneous computing platform are completed, hardware adaptability and computing compatibility are considered at the same time, and the overall operation efficiency of a system model is improved.

Description

technical field [0001] The invention relates to the realization of the software and hardware scheme of an image recognition system based on SNN, in particular to an image recognition software and hardware system based on SNN. Background technique [0002] Artificial intelligence has set off a new wave of development in various research fields. At the same time, as traditional industries are more closely connected with AI, AI empowerment has gradually become a technology hotspot in the context of the Internet of Everything. Among them, artificial neural network has been highly valued by researchers as a key part of realizing artificial intelligence. The development process of ANN can be roughly divided into three stages: from the initial simple computing model of the first-generation perceptron, to the second-generation fully-connected multi-layer continuous network model, and then to the current third-generation pulse transfer-based model A discrete network model of the fo...

AI Technical Summary

Problems solved by technology

[0003] In the field of image recognition with increasingly complex scenes and a large number of tasks, it has become increasingly difficult for traditional AI solutions to meet the real-time and energy-efficiency requirements of target recognition on the client side. Due to the transition of user equipment terminals, it is necessary to propose a new type of image self-classification pulse neural network model to reduce the hardware operation resource occupation and the overall complexity of the model

Images