Research And Application Of Lightweight Convolutional Neural Networks

With the rapid development of computer software and hardware and the arrival of the era of big data,deep learning has become a hot research topic,and has been widely used in various fields of life,such as face recognition,biomedical image recognition,public security criminal investigation,intelligent driving,etc.field.Convolutional Neural Network is one of the widely used deep learning algorithms.After continuous development,the accuracy of image recognition on large-scale data sets has exceeded the human average level and has become the top in the field of image recognition.Algorithm.CNN does not need to artificially extract image features during training,and can directly input pictures into the network to achieve end-to-end training and prediction,avoiding the complicated image preprocessing process of traditional recognition algorithms.CNN uses convolution kernel to extract image features.The structure is a multi-layer perceptron,which can extract more detailed image feature information in the image,and is still highly invariant to displacement,scaling,tilt and other deformation forms.Based on the convolutional neural network theory,this paper aims to design a lightweight and high recognition rate convolutional neural network and apply it to expression recognition.By studying the classical convolutional neural network structure and conducting network training and participating in network optimization,a network training model with fewer parameters and improved recognition rate is obtained.Finally,the model is used to construct the expression recognition system.The main research contents of this paper include the following aspects:(1)By comparing the classic AlexNet lightweight secondary improvement,comparing the performance of the model before and after the improvement on the dataset,verify the feasibility of the model lightweight and improve the recognition accuracy.The improved network uses parallel multi-scale convolution kernels to extract feature images.During the sampling process,the output dimensions are kept consistent,and then the output feature maps are integrated by cascading.The use of multiple convolution kernel sampling can increase the diversity of acquired features.The improved network uses a large number of 1×1 convolution kernels to reduce the feature map thickness,which is beneficial to the reduction of model parameters.The improved network was tested on the Caltech256 and 101_food datasets.The experimental results show that the improved network is reduced in size and recognition rate based on the original network.(2)In order to design a lightweight convolutional neural network,this paper designs a lightweight network SliceNet.When extracting features,SliceNet first divides the output feature map into two equal numbers.Each group uses different convolution kernels to operate.More sufficient information is extracted,and then the grouped diversity feature images are cascaded,and finally all the feature maps are integrated as the input of the next layer through the 1 × 1 convolution kernel.Compare the network performance with SliceNet's classification accuracy on Caltech256 and 101_food datasets using traditional NetNet.SliceNet increased the recognition rate from 50.1% to 52.2% on the dataset Caltech256 and increased the recognitionaccuracy from 66.3% to 68.9% on the dataset 101_food.(3)Based on SliceNet,a lightweight and high recognition rate network model ReduceNet is proposed.ReduceNet uses the Reduce module to lighten the convolutional neural network.The lightweight network still has a good recognition advantage when the model scale is greatly reduced.Based on the Reduce module,an improved network module ReduceV2 with the idea of residual is proposed.The lightweight network ReduceV2 Net was designed using the ReduceV2 module.ReduceV2 Net has enhanced performance in terms of recognition accuracy.(4)In order to apply the lightweight network in facial expression recognition,the deep learning framework Caffe is used to train on the facial expression dataset to obtain the facial expression recognition network model,and apply the recognition model to the real-time face in the video.Expression recognition.The integration of the network model results in an integrated model with better facial expression recognition performance and is applied to face and expression recognition in images.