Application Of Image Classification Technology Based On Convolution Neural Networks

With the rapidly development of deep learning in the field of machine learning and pattern recognition,the Artificial Intelligence technology has also made great progress.The convolutional neural network(CNN)has a simple structure,good robustness and strong adaptability.Currently,its excellent performance makes it widely used in the field such as machine vision applications,semantic segmentation,face recognition,gait analysis and pedestrian detection,and great success has been achieved,especially in image classification tasks.Image classification based on convolutional neural network has a higher accuracy than traditional methods.Hence,in this paper,human face images and finger knuckle print images are considered as the research object,and the image classification research is conducted by using convolutional neural network.The main works in this paper are summarized as follows:(1)Research on multi-task lightened face images model.This paper improves the model performance by data and algorithms.Through the preprocessing of alignment and enhance of training data,a face images network model was proposed.The characteristics learned by the network will have high discrimination and better network generalization performance by the increase of Center-loss layer.In this research,the training phase of the model is performed on the self-expanding database,and testing on the LFW(Labored Faces in the Wild)database.Experimental results show that the method used in this paper has higher recognition accuracy than the existing CNN model.In order to verify the universality of the face images model proposed in this paper,experiments were also performed on FER2013,SFEW2.0 expression database and Large Scale Asian Female Beauty Database(LSAFBD).The experimental results show that the proposed face images model is also suitable for face expression recognition and face beauty prediction,which has a certain improvement over the state-of-the-art methods.(2)Finger knuckle print recognition based on CNN.As a new object of biometric recognition,the finger knuckle print(FKP)is widely used in biometrics.In this paper,a convolutional neural network KnuckleNet is proposed for the identification of FKP images firstly.The region of interest(ROI)of the FKP image is extracted to reduce the amount of image processing calculation.Then,for small sample set,in order to solve the problem of overfitting,the original data is enhanced and a batch normalization layer is added after the network convolution layer.Finally,relevantexperimental results about model training and testing are given.Extensive experiments performed on the PolyU FKP database show that compared to traditional FKP recognition method,the proposed KnuckleNet network model can achieve better recognition results and has a good practical application value.(3)Face beauty prediction based on the residual network.The prediction of face beauty has become a new research topic and received more and more attention.To the best of author's knowledge,the application of deep learning methods for human face beauty is rare.Deep learning can learn features with higher discrimination and is also suitable for classification with smaller intraclass distances and larger interclass distances.In this paper,a method based on deep residual Network(ResNet)for predicting the beauty of human faces was proposed.By introducing residual optimization,the proposed method can solve the problem that the CNN gradient will disappear gradually as the layers go deeper.At the same time,for improving the prediction performance,the maximum feature graph activation function was introduced to make the network structure more compact,so as to effectively extract the effective features of human face image.The experimental results based on the Large Scale Asian Female Beauty Database(LSAFBD)show that the proposed method achieves a classification accuracy of 61.50% and 0.8620 person coefficient,which is superior to other CNN.