Research On Dense Convolutional Network For Image Classification

Image classification is an important and popular research topic in the field of computer vision.The development of deep learning has promoted the breakthrough progress of image classification.The exploration of image classification based on deep convolutional neural network(DCNN)has been relatively mature.Regarding image classification,improving the feature expression of DCNN is an important research to further improve performance.By analyzing and researching the densely connected convolutional neural network(Dense Net),we propose optimized networks,which combines Dense Net with soft attention mechanism to improve the feature expression ability of Dense Net and achieve more accurate image classification.First,the input of each convolutional layer is just a simple combination of its front feature maps in Dense Net.We propose Channel Feature Reweight Dense Net(CFRDense Net),which incorporates channel attention module into Dense Net.The channel attention module models the correlation between channels.In addition,convolution operation is local.We introduce a feature point attention module in the CFR-Dense Net constructing Channel-wise and Feature-points Reweights Dense Net(CAPR-Dense Net).The CAPR-Dense Net adaptively calibrates the channel feature response and explicitly models the interdependence between point-wise features.Through image classification experiments on the CIFAR dataset,we prove the effectiveness of the CAPR-Dense Net.Secondly,the Dense Net does not fully consider the inter-layer feature correlation.We integrate the inter-layer feature attention module into Dense Net to build an Inter-Layer Feature Reweighted Dense Net(ILFR-Dense Net).In addition,we use Ensemble method to combine the ILFR-Dense Net with the CAPR-Dense Net.We construct Multiple Feature Reweight Dense Net(MFR-Dense Net),which takes full advantage of ILFR-Dense Net and CAPR-Dense Net to further improve the feature expression.In order to verify the effectiveness of the method,we perform image classification experiments on the CIFAR dataset.The experiment found that the parameter of the 100-layer MFR-Dense Net is only14.2M,but the test error rate on CIFAR-10 arrives 3.57%,and the test error rate on CIFAR-100 arrives 18.27%.The image classification performance of the MFR-Dense Net on the CIFAR-10 and CIFAR-100 datasets is better than most current previous methods.Finally,MFR-Dense Net has many shortcomings: it cannot be trained end-to-end and the training process is cumbersome;parameters and calculation are relatively large;training and testing take longer.Therefore,we propose a Dual Feature Reweight Dense Net(DFR-Dense Net)which can be trained end-to-end.Through experiments on CIFAR-10/100,MORPH and Adience(high-resolution datasets for face age),it was found that compared with the MFR-Dense Net,the parameters of DFR-Dense Net was reduced by half,and the test time was reduced by about 61%.And end-to-end DFR-Dense Net can enhance the learning ability and improve the accuracy of image classification.The DFR-Dense Net has certain adaptability and practicability to different datasets.This paper proposes optimization networks for Dense Net.Through experiments on CIFAR,MORPH and Adience,it is found that the method proposed in this paper can enhance the feature expression ability and effectively improve the accuracy of image classification.