A Study Of Image Recognition Techniques Based On Convolutional Neural Networks

Due to its strong representation learning ability,convolutional neural networks(CNNs)have attracted considerable attention and made rapid development in the fields of image recognition,etc.For image recognition,existing studies mainly improve the performance and efficiency of the related tasks by building general neural networks or special networks for fixed problems.However,the existing CNNs still have the following shortcomings:1)they usually need a large amount of training samples for enhancing the deep feature learning,and in the case that only a small amount of training data containing noise is available,both the robustness and recognition performance of existing models will be degraded;2)since most existing CNNs utilize the standard convolutions or the standard convolution based related modules for representation learning,however the inefficiency of the standard convolution operation will directly make the related modules and even whole networks inefficient.In addition,the traditional pooling operation is not learnable,which may lose important feature information;3)they usually ignore the usage of hierarchical features or only pay attention to local features,the global feature learning ability is not sufficient,and moreover the feature information fusion ability from different layers of the network is relatively lacking.To this end,this thesis will mainly propose three innovative solutions to overcome the above drawbacks,and verify the effectiveness of the new models through public image datasets.The main contributions of this thesis include(1)To improve the feature learning ability,robustness and recognition performance of the CNNs in the case of a small amount of training data,a new representation learning algorithm based on the low-rank sparse convolutional features is proposed.This algorithm further optimizes the deep features extracted by CNNs by fully excavating the low-rank sparse features of the deep convolutional features and denoising them at the same time.In addition,the local relationship among features will be preserved in the optimizations at the same time.Compared with the original method,our new model can effectively enhance the feature learning ability and strengthen the image recognition ability of traditional CNNs in the case of a small amount of training data that possibly cantains noise.(2)An end-to-end folly-convolutional intensive feature flow network is proposed,which can enhance the representation learning ability and make full use the hierarchical features.Besides,our model also solves the unlearnable issue of the pooling operation and the fragile propeties to lose features.In our framework,a new intensive block is designed,which can enhance the fluidity and coupling of convolution features in the local and global learning scenarios,thus enhancing the utilization of local and global features to a certain extent.In addition,to alleviate the low efficiency issue of traditional convolution operation,the depthwise separable convolution is used in our new network to improve the efficiency of the fully-convolutional network.(3)To improve the efficiency of dense block in DenseNet,a new fast dense block(FDB)is proposed.The idea of FDB is also introduced into the residual dense block in RDN,which leads to a new block called fast residual dense block(f-RDB).By this way,one can not only retain the learning ability of local features of RDB,but also reduce the computational cost and improve the efficiency at the same time.For enhancing the learning ability of global features,a new fast dense residual network is further proposed,in which a global dense block(GDB)that can fully use and learn global feature information is constructed.After the module in the network,a down-sampling block is also included,which has gradually increasing number of convolutional channels to extract more informative deep features,while reducing the feature size.Finally,our new network can make full use of both the local and global feature information to enhance the recognition performance.The effectiveness of the proposed feature learning algorithms has been verified by extensive experiments on public image databases.