SqueezeNet based handwritten digital recognition method

Abstract

The invention discloses a SqueezeNet based handwritten digital recognition method. The SqueezeNet based handwritten digital recognition method specifically comprises: (1) pre-processing recognized handwritten digital samples; (2) constructing a SqueezeNet convolutional neural network model; (3) training the SqueezeNet network model, and testing the handwritten digital samples; and (4) obtaining arecognition result. Compared with the traditional AlexNet, the SqueezeNet network model in the invention effectively reduces the total number of network parameters, and improves the accuracy of handwritten digital recognition, has the basis of model compression, and can be better applied to mobile devices, distributed training and embedded hardware.

Description

technical field [0001] The present invention designs the field of image recognition and deep learning, specifically a handwritten digit recognition method based on SqueezeNet. Background technique [0002] With the rapid development of deep learning, convolutional neural network (CNN) has been widely used in image recognition fields such as handwritten digit recognition, and has achieved excellent recognition results. However, the number of parameters that need to be trained in the network is too large, resulting in high recognition accuracy of CNN handwritten digits, but the required training time is long, which cannot be effectively applied in distributed training (Forrest N.Iandola et al. 2016 proposed The communication overhead of distributed data parallel training is proportional to the number of parameters in the model), and it is difficult to import the model to mobile devices and deploy it on FPGA and some embedded hardware. Jiantao Qiu et al. proposed that the model...

AI Technical Summary

Problems solved by technology

However, the number of parameters that need to be trained in the network is too large, resulting in high recognition accuracy of CNN handwritten digits, but the required training time is long, which cannot be effectively applied in distributed training (Forrest N.Iandola et al. 2016 proposed The communication overhead of distributed data parallel training is proportional to the number of parameters in the model), and it is difficult to import the model to mobile devices and deploy it on FPGA and some embedded hardware. Jiantao Qiu et al. proposed that the model should be as small as possible The model of the model is stored on the FPGA of less than 10M, and should not let the storage bandwidth limit its storage

The AlexNet structure proposed in 2012, although the recognition accuracy of ImageNet has been improved to a certain extent, and the generalization ability of the network has also been enhanced to a certain extent. At the same time, multiple GPUs are used to accelerate the calculation of the network. The parameters have not been effectively reduced, and the network model is still too large

[0003] Nowadays, digital recognition technology has been widely used in our lives, such as license plate recognition, ID card recognition, bank bill recognition, etc., and has a good recognition effect, but due to too many network parameters and too large a network model, The overhead required by the network model is still large, and the application on hardware devices is also greatly restricted

Images