Research On Application Of Deep Learning Models For Feature Representation And Classification

With its layered end-to-end learning network frameworks and local to global feature representation technology,deep learning is becoming a research hotspot in the fields of pattern recognition and computer vision,and has been successfully applied to object detection and face recognition.This paper has conducted on small size dataset learning,discriminant loss function design,and few-shot learning in existing deep learning technologies,and has proposed some improved algorithms and model structures.It has been successfully applied in face recognition and target classification.The main work of this article is as follows:1.Deep Convolutional Neural Network with Dilated Convolution Using Small Size DatasetAiming at the problem of traditional deep convolutional neural networks facing small size datasets(non-overcomplete datasets),they will fall into serious overfitting problems,and a new deep dilated convolutional neural network framework is designed.The framework first uses dilated convolution to replace the traditional convolution to increase the receptive field of the network;Then reduce the parameter size of the fully connected layer by using a global average pooling layer instead of the first two fully connected layers;Finally,batch normalization is added to the convolutional layer to weaken the Internal Covariate Shift,thereby speeding up the network training speed.The improved network has better classification capabilities for small size datasets,effectively reducing the network's overfitting problem.2.Double Additive Margin Softmax Loss for Face RecognitionAiming at the defect of insufficient discriminative feature extraction ability in Large-Margin Softmax Loss and Additive Margin Softmax Loss,we propose the Double Additive Margin Softmax Loss.The algorithm adds an additive margin m to both the intra-class angle variable and the inter-class angle variable to enhance the intra-class compactness and inter-class difference of features.Extensive experimental evaluation of several recent state-of-the-art softmax loss functions are conducted on the relevant face recognition benchmarks,CASIA-Webface,LFW,CALFW,CPLFW,and CFP-FP.We show that the proposed loss function consistently outperforms the state-of-the-art3.Collaborative-Labeling Few-shot Learning Based on Graph Neural NetworkAiming at the problems of lack of nonlinear identification ability,insufficient feature receptive field and lack of global features in the Edge-Labeling Graph Neural Network,we propose the Collaborative-Labeling Few-shot Learning Based on Graph Convolutional Neural Network.The algorithm first uses an expanded convolution filter to replace the traditional convolution filter of the last three convolution blocks in the embedded network to increase the feature receptive field;Then use the encoder in the convolutional auto-encoder to replace the first convolution block in the embedded network to enhance the feature extraction capability of the network;Finally,we increase the input feature size of the metric network by adding blank pixels and replace the 1×1 convolution filter with a 3×3 convolution filter to increase the receptive field of the network and increase the global features of the network.The experimental results on the mini-ImageNet and tiered-ImageNet also verify that our network is superior to Edge-Labeling Graph Neural Network.