Research On Image Recognition Based On Optimal Convolution Neural Network

With the continuous improvement of computer performance and the continuous advancement of science and technology,deep learning has become a new research hotspot and has achieved breakthrough research results at the age of big data,and in image processing,autopilot,speech processing,natural language processing and other fields have been widely used.The convolution neural network is a commonly used deep learning algorithm.It has great advantages in image processing,especially in image recognition,and it shows very good performance.It can be said that it is one of the best image recognition methods at present.Compared with the traditional image recognition method,the Convolution Neural Network can directly use the image as a network input,through the layer-by-layer operation and transfer,the distinguishable image features can be automatically extracted,and finally through the classifier completes the recognition process.The entire process avoids excessive preprocessing and manual selection features in traditional methods and completes end-to-end image recognition.In this paper,we combine the basic theory of convolution neural network for image recognition,and aim is to design a network that can improve the recognition rate and reduce the parameters.By combining the classical structure of convolution neural networks in recent years,and by continuously tuning and participating in the optimization of the network so that we get good results.The model achieves the goal of improving network accuracy and reducing parameters.The main research contents and innovations of this article are as follows:(1)Introduced the basic concepts of convolution neural networks,mainly include the working principle,back propagation algorithm and overall structure;Introduced the Caffe framework for deep learning and briefly described in data preprocessing such as normalization of scales,mean denoising,and data enhancement.(2)Combined with the development of current convolution neural networks,MyNet network is designed as a reference network for image recognition.In order to improve the recognition rate of the network,it improved the reference network and designed a multi-scale and parallel crossover network.The multi-scale convolution network uses convolution kernels ofdifferent scales to extract features,and then merges the features extracted from different scales;the parallel crossover network uses two channels to feature extraction respectively,and the two channels of feature maps are cross-fused before the feature size is reduced.Finally,the three networks were tested on the cifar10 and cifar100 data sets,and the results show that the improved network can increase the recognition accuracy of 2.11%,2.12% and 1.19%,3.32%respectively on the two data sets.(3)In order to reduce the network parameters,the overall structure and related parameters of the AlexNet network and SqueezeNet network are first analyzed,and then combined with the idea of the SqueezeNet network to improve on the AlexNet network,and designed a lightweight network model of cascading and merging.The model is mainly to replace a convolution layer with two convolution layers and reduce the parameters by compressing the number of input channels.Finally,we do an experiment on Food-101 and GTSRB data sets,and the results show that the improved network reduces the network parameters to a certain extent and it can increase the accuracy of 5.18%,6% and 1.21%,1.42% respectively on the two data sets.(4)In order to further reduce the parameters of the network and continue to improve the recognition rate of the model,two ultra-lightweight convolution neurons were designed by analyzing the techniques of some subtraction parameters,such as bottleneck module structure and grouped convolution structure,and combining the idea of residual network,then explained the characteristics of the model.we elaborated the ultra-lightweight network mainly include the basic modules,overall structure and specific parameter settings.Finally,a better model was obtained by continuously optimizing the network and experimenting on the Food-101 and GTSRB data sets.The results showed that the ultra-lightweight network achieves a large number of parameters to reduce and significantly improve the network recognition rate.Compared with AlexNet,the size of the network model is reduced to about 3.3% and 2.8% respectively on the two data sets,and the accuracy is increased by 13% and 2% respectively.