The Research And Applications Of Gaussianscale Space Convolutional Neural Network

Convolutional neural network is a partially connected feed-forward neural network model.A representative CNN model contains convolutional layer,pooling layer and fully-connected layer.Using convolution computation,the convolutional layer can extract feature step by step and layer by layer.Besides,the pooling layer can reduce the dimension of features,and the fully-connected layer can establish a mapping relation between the input images and target information.Through proper training,convolutional neural network can classify massive image data,and the performance of CNN is outstanding.In recent years,deep learning technology is receiving increasing attention.The theory and the applications of convolutional neural networks have made a continuous development.Traditional CNN is robust for images'distortion,rotation,and scaling to a limited extent.But when the test image is far different from the training data in the training set,such as different in size or scale,the robustness of CNN is weak.And CNN models may occur scale mismatch.In this case,the classification accuracy of CNN model is not quite high.In order to solve this problem,some domestic and foreign researchers proposed many improved models to enhance CNN's multi-scale processing capacity.In this thesis,we present a new CNN model which is based on Gaussian Scale Space.We call it Gaussian Scale Space-CNN?GSS-CNN?.The basic design idea of GSS-CNN is to change the way of feeding single size or single scale training data to CNN model,and design a multi-channel CNN model which can read several images at the same time.These images come from Image Pyramid.Thus,the larger image have smaller smoothing factor.In this model,the convolutional neural network can extract images'features from multiple sizes and multiple scales.Thus,the CNN's ability of adapting to input images can be enhanced.In order to testing the performance,we designed three GSS-CNN models,i.e.GSS-CNN_II,GSS-CNN_?andGSS-CNN_IV,and make some experiments.The result shows that GSS-CNN has higher classification accuracy when we choose the number of channel properly.The classification accuracy ofGSS-CNN_?is 2.276 percentage higher than traditional CNN models.Meanwhile,GSS-CNN has higher tolerability of scale changing.If we change the scale of test images,GSS-CNN's performance is always better than CNN.We also study the effect of the number of channels on the performance.The experiment proves that the three-channel GSS-CNN is a good choice,since it has the highest classification accuracy.In order to further improve the performance of GSS-CNN,this thesis considers the Stacking ensemble strategy in ensemble learning,and proposes a GSS-CNN model based on Stacking ensemble strategy?Stacking GSS-CNN?.We redesign the structure of GSS-CNN,so that every channel in this model has a weighting factor.Compared with GSS-CNN,Stacking GSS-CNN achieves a higher classification accuracy.The experiment shows that when theGSS-CNN_?combined with MLR meta-learner,this model's classification accuracy which is based on MNIST can be 92.16%.This result is 1.26%higher thanGSS-CNN_?.Obviously,the performance of Stacking GSS-CNN is better.