Convolution nerve network acceleration method based on pre-deciding and system

Abstract

The invention discloses a convolution nerve network acceleration method based on pre-deciding, comprising steps of obtaining a calculation price of each characteristic point and dependency relation between the characteristic points according to the characteristics that CNN forward propagation and each layer calculation mode, establishing a characteristic point selection model on the basis of the calculation price of each characteristic point, the dependency relation between the characteristic points and the classification capability of the characteristic point, configuring a series of incremental characteristic point quantities according to the structure characteristic points, using the characteristic point selection model to perform one-by-one optimization on the series of incrementing characteristic point number according to a training data set which is established in advance to obtain linearity classifiers which are chosen by and corresponding to the series of characteristic points, and constituting all linearity classifiers into cascade classifiers according to the linearity classifiers which are chosen by and corresponding to the series of characteristic points and the preset threshold. The invention fully utilizes the CNN characteristic redundancy, the multilayer characteristic calculation cost variation and multistage characteristic discrimination capability to realize the acceleration of the CNN on the binary classification and the specific object detection.

Description

technical field [0001] The present invention relates to the field of computer vision and depth technology, in particular to a pre-decision-based convolutional neural network acceleration method and system. Background technique [0002] At present, CNN is widely used in the field of computer vision, and its performance is also very superior. It has achieved leading performance on problems such as image classification and object detection. However, the high computational complexity of CNN has limited its use in application scenarios with high real-time requirements, such as traffic monitoring for pedestrians and vehicles; it is used on devices with low computing power. There are also great limitations in applications, such as mobile devices. [0003] The existing CNN acceleration methods are mainly divided into the following five aspects: 1. Lower the rank of the CNN convolution kernel; Convolution operations in the Fourier frequency domain; 4. Acceleration methods based on...

AI Technical Summary

Problems solved by technology

[0004] However, existing acceleration methods still have flaws and have not completely solved the speed bottleneck of CNN in applications

In addition, the existing acceleration methods have not considered the redundancy of features extracted by CNN and the possibility of using its multi-layer features for multi-level decision-making.

Images