Deep Convolutional Neural Network With Multi-scale Sharing And Subpatch Cascading For Object Recognition

Visual object recognition is an important task in the field of artificial intelligence.It is the foundation for machines to sense the surrounding environment through cameras.Visual object recognition is aimed at recognizing specific objects in the given images or videos and generally consists of three procedures: region proposal,feature extraction,and target classification.Traditional object recognition methods typically rely on hand-crafted features whereas deep neural network based methods are able to automatically learn features via end-to-end learning on training dataset.In recent years,deep neural network based methods have developed rapidly and surpassed traditional methods by a large margin.Therefore,they are widely applied in a variety of fields,such as security monitoring,information security,intelligent transportation,self-driving vehicles,and human-computer interaction.This dissertation proposes to share parameters across multi-scale deep convolutional neural network models to speed up object recognition.In addition,this dissertation proposes a new network structure composed of cascaded subpatch filters to improve the recognition accuracy.On the basis of above research,this dissertation presents one elastic nonlinear activation function to further improve the recognition accuracy.All in all,this dissertation achieves three meaningful results listed below.1.On the aspect of transferring network features,this dissertation proposes to share multi-scale model parameters of deep convolutional neural network.This method can reduce the computation consumed on calculating features of different scales,thus speeding up the recognition process.Experimental results show that the proposed method effectively accelerates object recognition with negligible loss of recognition accuracy.2.On the aspect of designing network structure,this dissertation proposes a new convolutional structure based on cascaded subpatch filters.This method is able to strengthen the representation ability of the extracted features by imposing a group of subpatch filters on the local image patch,thus strengthening the recognition ability of the whole network.Experimental results demonstrate that the proposed method can improve the recognition performance without increasing the computation complexity.3.On the aspect of network activation function,this dissertation proposes one elastic nonlinear activation function.Activation function is generally one nonlinear function and it is the foundation for the network to deal with nonlinear data.The proposed elastic nonlinear activation function not only can impose non-linear transformation on the neurons but also can be seen as one stochastic regularization technique.This helps improve the generalization ability of the network.Experimental results demonstrate that the proposed method can improve the recognition performance without bringing in extra parameters.