Research On Image Classification Application Based On Transferring Convolutional Neural Network

Image recognition is an important component in the field of computer vision.It enables the computer to automatically recognize and process images by identifying the different categories of objects in the image.Among different image recognition tasks,facial expression recognition(FER)and Synthetic Aperture Radar(SAR)target recognition have been used in human-computer interaction,medical treatment,cardiac consultation,and military,surveying,and resource exploration.They have important application value,and more and more scholars and scientific research institutions are focusing on facial expression recognition and synthetic aperture radar target recognition.The image recognition process is generally divided into three parts,namely image preprocessing,image target feature extraction and image target recognition.For facial expression recognition tasks,this paper mainly studies the problem of insufficient training data in this task and how to effectively use web face expression data,and this paper also studied the transfer learning mechanism for facial expression recognition.For SAR image target recognition tasks,this paper focuses on how different convolutional neural networks could affect the SAR image recognition.The influence of different CNN structure on the SAR image target recognition performance was studied.Besides,this paper studied the recognition influence Inception model could bring.In response to the above questions,this paper mainly did three works:(1)To solve the problem of insufficient training data in face expression tasks,this paper utilized the transfer learning method to migrate the deep convolution neural network trained on face recognition tasks to face expression recognition tasks,in order to enhance the recognition ability of the network,Double activation layer network structure and Softmax-MSE loss function are applied to the convolutional neural network.Experiments show that the method proposed in this paper can greatly improve the model recognition ability,and achieved 61.59% and 47.23% on the FER 2013 database and the SFEW 2.0 database,respectively.(2)In the Internet and various types of social software,a large amount of facial expression image data can be utilized to boost the performance of facial expression recognition tasks.However,the problem of noisy label is widespread in network face expression data,which brings a great challenge to web-based face expression data.In order to better solve the problem of mislabeling in network face expression data,this paper adopts a method of emoticon recognition model training based on web network facial expression data based on active learning and transfer learning.At the same time,this paper also explores the two-stage transfer learning process.This paper carries out experiments on FER2013 database and SFEW2.0 database and web expression dataset,which is namely the Web EXP dataset,and verifies the effectiveness of this method.(3)In order to study the effects of different convolutional neural network structures on SAR image target recognition performance,this paper first performs database enhancement on the MSTAR database.By rotating the image in the database clockwise at every one degree,the original database is expanded 360 times.Then,the influence of convolutional neural network on SAR image target recognition is studied from different convolution kernel size,different number of convolution layers and whether there is ReLU activation function.Based on above exploration,a two-layer convolutional neural network model was constructed and the two-layer convolutional neural network is trained on the MSTAR database after the data enhancement processing.Furthermore,a Inception model based CNN network,which was Inception-SAR,was also proposed to boost the performance of SAR recognition.The experimental results show that The neural network has achieved good experimental results in the three types of SAR target classification experiments.