Several Application Research Of Image Recognition Based On Deep Convolution Network

Image recognition automatically recognizes different kinds of target images by extracting the feature information contained in the image.It's one of the important parts in the field of machine vision.How to effectively process,analyze and apply image data has always been a hot topic in the field of vision.Image recognition includes image acquisition,image pre-processing,image feature extraction and image recognition four steps.Among them,how to extract image features with strong expressive ability and robustness remains to be solved.Deep convolution network has strong non-linear fitting ability,which makes it widely used in many fields,such as face recognition,biometric recognition,SAR target recognition,person re-identification and so on.Compared with traditional image recognition methods,deep convolution network reduces the interference part of artificial feature design,and achieves better performance through reasonable structure design.However,for a specific image recognition task,how to design an effective convolution structure remains to be explored.In this paper,image recognition based on deep convolution network for facial beauty,finger-knuckle-print and SAR images are studied.The research mainly contents following three aspects:(1)Unconstrained facial beauty prediction based on BeautyNet.To enhance the expression of facial beauty features,this paper constructs a multi-scale deep convolution network BeautyNet.BeautyNet integrates multi-resolution and multi-scale facial beauty features to enhance the expression of features.In order to alleviate the computational complexity introduced by the multi-scale structure,the MFM activation function is used as a non-linear element.The MFM activation function could reduce the computational complexity of the network and accelerate the convergence speed of the CNN.At the same time,to alleviate the over-fitting problem of fewshot learning,this paper uses transfer learning strategy to transfer face information from CASIA-WebFace database to facial beauty prediction task.The experimental results show that the combination of BeautyNet and transfer learning strategy could greatly improve the performance of facial beauty prediction.BeautyNet achieves 67.48% classification accuracy on a large scale facial beauty database,LSFBD.(2)Finger knuckle print recognition based on SlimNet.A SlimNet network is designed for finger knuckle print recognition.Among them,the network uses small-scale convolution block SlimBlock to construct the main body of SlimNet,which reduces the computational complexity of the model.At the same time,in order to improve the quality of FKP image,we extracted the ROI of Finger-Knuckle-Print database of Hong Kong Polytechnic University before SlimNet training.Concretely,the edge detection and rectangular window traversal search method were used to detection and crop the ROI region of FKP image.To alleviate the over-fitting phenomenon of SlimNet,360-fold rotation enhancement method is used to expanded data.The experimental results show that the combination of SlimNet network,extracted the ROI of FKP images and data enhancement of rotation,the SlimNet could achieve the optimal FKP recognition performance.For left index finger,left middle finger,right index finger and right middle finger four database,SlimNet achieved 97.65%,97.79%,97.31% and 97.96% recognition accuracy,respectively.(3)SAR target recognition based on MiniNet.A simple CNN structure is designed to achieve efficient SAR target recognition,which only contains two convolution layers.In order to reduce the environmental noise interfere of the original database of SAR,the centroid method is used to extract ROI from SAR image before target recognition.In order to effectively alleviate the problem of insufficient data for SAR target recognition,the ROI image is enhanced 360-fold though rotate SAR images.In terms of experiment settings,the effects of training epoch,convolution kernel size and convolution layers on SAR target recognition accuracy are explored.By optimizing the parameters of the model,MiniNet achieves 97.29% recognition accuracy on the MSTAR database.