Research On Image Classification Based On Improved Capsule Neural Network

For three-dimensional images such as magnetic resonance images(MRI),hyperspectral images(HSI)and video sequence images,Convolutional neural networks(CNN)has to rely on a more complex network structure when classifying images,and it is difficult to improve the accuracy.In addition,MRI and HSI have complex structure,less available data and rich spatial information.Compared with the video sequence image,there is no temporality between channels.CNN has the problems of few training samples and insufficient feature extraction for this kind of image classification.At present,deep neural networks based on multi-scale methods have shown excellent capabilities in image classification;at the same time,the capsule neural network has achieved extremely high accuracy in the field of two-dimensional image recognition due to its ability to retain spatial information,but there are also problem of very large amount of parameters.In order to further study the effect of the capsule network on the classification of complex images and the potential of optimizing parameters,a multi-scale capsule network optimized with reducing parameters is proposed with MRI and HSI as the research object.The main work includes: First,combine the capsule network with the two-branch neural network,randomly sample the input images,and extract multi-scale information using the two-channel capsule network.It can not only solve the local effects and spatial information loss caused by the pooling layer,but also fully extract the multi-scale information from the input.Second,Dual-branch 2D-Conv Caps Net(Du B-2CCN)is proposed for the rich spatial information of brain tumor MRI,and Du B-Conv Caps Net-MRF is proposed for the spectral-spatial information of hyperspectral images.Third,according to the difference between the two architectures,1D and2 D constraint windows are proposed to reduce the number of capsules in the corresponding dimensional channel,which uses the capsule vector group as the calculation unit to perform convolution operations.This reduces the number of parameters and the calculation complexity of the capsule network.The experiment uses a pixel-by-pixel classification method.The experiment is compared with the current advanced algorithm on the Bra TS2015 data set,and it is better than other methods in most evaluation indexes.Comparing the model with the second performance in each index,the Dice score is 3.4% and 6.9% higher in the complete tumor and the tumor core area.The parameter-reducing optimization method of the 2D constraint window proposed in this paper reduces the parameter amount of the double-branch capsule neural network by 12.6% and the training time by 42.1%,and increases the computational intensity by 26.9%.The proposed method Du B-Conv Caps Net-MRF is 0.21% and 0.27% higher in OA indicators than the models with the second performance respectively on the hyperspectral datasets of Indian Pines and Pavia University.The proposed parameter reduction optimization method reduces the training time of the capsule network method by 37.5%.In summary,the proposed capsule network for the feature extraction method is suitable for two different types of high-dimensional data classification,which makes the model have translational variability and no longer depends on a large number of data sets;the capsule network with the improvement of additional parameter optimization strategies reduces the pressure of dynamic routing,greatly reduces the amount of parameters,and can prevent overfitting to a certain extent.