Research On Multispectral Image Segmentation Based On Convolutional Neural Network

As a key technology in image processing,image segmentation is widely used in object coloring,stereoscopic scene analysis,medical image processing,etc.Since multispectral image data have rich spatial information and spectral information,traditional threshold-and region-based image segmentation methods are difficult to effectively segment images.In recent years,with the advancement of deep learning technology and the improvement of data processing capabilities,the image segmentation based on convolutional neural network(CNN)is the trend of future development.In this thesis,we use deep learning and tensor decomposition theory to study the segmentation problem of multispectral images.The specific work is as follows:Firstly,based on the rich but redundant band characteristics of multispectral images,an adaptive band selection mechanism based on CNN is proposed.The main idea of this mechanism is to extract the characteristics of each band in the multispectral image by CNN.Then,the three best band subsets are selected according to the characteristics,which achieves the conversion of multispectral images to three-spectral images.Then,a full convolution segmentation network that includes up-sampling and down-sampling is designed to connect to the adaptive band selection network.Finally,the whole network can realize the end-to-end pixel class prediction of multispectral images.Secondly,a multi-spectral image segmentation scheme based on low-rank tensor decomposition is designed to reduce the number of parameters in the above network.The VGG classification model by Tucker decomposition is adopted as a parameter initialization.Then,Tucker decomposition is performed on the convolution kernel of the whole network.Finally,the parameters of the network are updated by training again.This model can reduce the memory usage of the device and the execution time while realizing multispectral image segmentation.Finally,the proposed schemes are validated by FV and Portrait data sets on the Python platform based on TensorFlow deep learning framework.In order to optimize the segmentation edge effect,the super-resolution reconstruction guide filter layers are added to the proposed schemes in the experiment.The experiment results show that the proposed schemes can achieve the pixel accuracy rate of about 90%and the average class accuracy rate of 80%on FV and Portrait data sets.Furthermore,the number of parameters in the scheme based on the low-rank tensor decomposition is reduced by about 70%while the performance of the segmentation is not significantly reduced.