Method for accelerative compression of deep convolutional neural networks for handwritten Chinese character recognition

Abstract

The invention discloses a method for accelerative compression of deep convolutional neural networks for handwritten Chinese character recognition. The method comprises the following steps of constructing and training a deep convolutional neural network for handwritten Chinese character recognition; carrying out decomposition training on convolutional layers of the deep convolutional neural network layer by layer by adoption of a low-rank decomposition strategy, so as to decrease the calculated amount; removing redundancy connection of the convolutional layers and a full-connection layer of the deep convolutional neural network by adoption of a network pruning strategy, so as to decrease the storage amount; compiling a forward code of the deep convolutional neural network. Compared with the prior art, the method disclosed by the invention has the advantages that the low-rank decomposition strategy for the convolutional layers and the pruning compression strategy for the whole deep convolutional neural network are adopted at the same time, so that the calculated amount and storage amount of the deep convolutional neural network are greatly decreased; the low-rank decomposition strategy and the pruning strategy for the deep convolutional neural network are effectively fused, so that an effective accelerative compression effect of the whole deep convolutional neural network is achieved.

Description

technical field [0001] The invention relates to the technical fields of pattern recognition and artificial intelligence, in particular to an accelerated compression method of a deep convolutional neural network for handwritten Chinese character recognition. Background technique [0002] Chinese characters originated from pictographs, which have a history of thousands of years, and are still the most widely used characters in the world. The large number of writers means that they have many writing styles, a large number of categories, and the existence of similar characters, which leads to out-of-the-box Mobile handwritten Chinese character recognition has always been a difficult problem in the field of pattern recognition and artificial intelligence. At present, offline handwritten Chinese character recognition has been widely used in many fields such as ancient document scanning, handwritten bill recognition, scanners, and text recognition in natural scenes. [0003] Now b...

AI Technical Summary

Problems solved by technology

[0004] However, in order to use convolutional neural networks to achieve a better result, people often directly make the network bigger and deeper, or even train multiple networks, and then integrate them. Although this approach can slightly improve the recognition rate of the network, However, the amount of calculation and parameter storage it introduces will increase dramatically

As a result, offline handwritten Chinese character recognition cannot be applied to real life on a large scale, especially in mobile devices and embedded settings.

Images