Visualization Algorithm And Application Of Deep Neural Network

At present,the great success of deep learning in the field of computer vision has made it one of the most influential research directions.Many studies have improved the deep network and applied it to various fields.However,improvements to deep networks is mainly to increase the depth of the network or embed more and more complex models in the network.Although this improved idea can improve the performance of the network to a certain extent,it also leads to the network becoming more and more bloated,the greater the difficulty of network training and the increase in the number of training data sets.At the same time,this kind of network structure improvement does not involve the internal processing algorithms of deep networks,which is not conducive to the theoretical or algorithmic breakthrough of deep learning.Therefore,we need to analyze the internal processing mechanism of deep networks,so according to different The problem is improved by the internal algorithm of the deep network.Firstly,the depth network is visualized layer by layer with the maximum activation visualization algorithm,so as to further understand the main role of each layer and each convolution kernel of the depth network.Based on the research and analysis of the internal processing mechanism of the deep network,the visualization-based deep network compression algorithm,the image super-resolution reconstruction algorithm based on the deconvolution neural network,and the spatial-temporal fusion algorithm of remote sensing images based on the dense connection network are studied.Details are as follows:1.Depth-based network compression algorithm based on visualization.Firstly,the depth network is visualized layer by layer using a classical AM algorithm.By analyzing the visualization results,we find that different convolution kernels extract different features,and there are many convolutional kernels that extract the same feature or invalid information.These redundant convolution kernels waste a lot of computing resources and storage resources.In response to this phenomenon,we propose a convolutional kernel pruning strategy based on the visualization results.For a single convolutional layer,we propose a optimization model to prune convolutional kernels.When the convolution kernel is deleted,it is required to obtain as much information as possible in the convolution layer.And we use the greedy algorithm to solve the optimization model.Then,we extended the single-layer pruning strategy to the entire network and compressed the deep network as a whole.Compared with the general deep network compression algorithm based on convolution kernel importance index,our proposed visualization-based compression algorithm can not only delete invalid convolution kernels,but also remove convolution kernels with similar functions.The comparison experiment shows that the proposed algorithm has better compression performance.2.Deconvolutional neural network for image super-resolution.There is a fundamental difference between image classification problem and image reconstruction problem.The great success of convolution operation on image classification problem does not mean that it is the best choice for image reconstruction problem.The convolution operation extracts the most efficient feature for categorizing images from a bunch of features,and is a process in which information is changed from more to less.Image reconstruction problems require more high-resolution details to be recovered from a small amount of low-resolution information.Information is a process in which features change from less to more.The deconvolution operation can be regarded as the inverse operation or the transposition operation of the convolution operation,and can recover more information from a small amount of information.Therefore,the principle of the deconvolution operation is more consistent with the mechanism of the image reconstruction problem.Based on this,we propose a deconvolutional neural network for image super-resolution.Most of the networks used for image reconstruction choose the mean square error(MSE)as the loss function of the network,just like the image classification network.However,the MSE loss function will cause the recovered image to be too smooth due to its too smooth characteristics.Therefore,in order to avoid this defect of MSE,we regard the image as a distribution of pixels,and introduce the Kullback–Leibler(KL)divergence into the loss function of the network.Through comparative experiments,although the proposed FDNN network has only 10 deconvolution layers,the reconstruction effect is better than the 20-or even 30-layer deep convolutional networks.3.Spatial-temporal fusion algorithm for remote sensing images based on multi-input dense neural network.The general network for image superresolution reconstruction is single-input and single-output,and the remotetime image spatio-temporal fusion problem contains sequence images.In order to make full use of the information of adjacent-time images,we propose a multi-dense neural network(MDN).The network uses two lowresolution MODIS remote sensing images at adjacent times as input to obtain an excessive high-resolution remote sensing image.Considering that the space-time fusion problem of remote sensing images is only missing high-resolution images at certain moments,both high-and low-resolution images before and after the missing moments are available,and the MDN network only uses the information of low-resolution images.Therefore,we propose a new error linear interpolation(ELI)algorithm,which fuses the two excessive images learned by MDN with two high-resolution images before and after the known missing moments to obtain the final Landsat image with high resolution at the middle missing moment.Through contrast experiments,the fusion algorithm proposed in this paper is superior to the traditional sparse representation-based fusion algorithm and the general single-input single-output image super-resolution reconstruction network in quantitative comparison and visual contrast.