Polarimetric Imaging Target Classification Methods And Experiments Based On Convolutional Neural Network

Polarization is one of the basic properties of light.When sunlight incidents on different target,it will produce polarized light in different states.The polarization state of the reflected light is closely related to the material,shape,roughness and other factors of the target.Therefore,different objects have different polarization characteristics,especially between the the artificial target and the natural backgrounds.With the development of optical imaging technology and the increase in the difficulty of target classification tasks in clutter environments,more information can be obtained by the polarimetric imaging technology compared with the traditional one,so it has been widely used in various fields.With the development of deep learning technology,the current main target classification solution is implemented by using convolutional neural networks,which is characterized by data-driven feature learning.Compared with the traditional methods,the convolutional neural network can automatically express more complex mathematical models and has stronger generalization ability.However,the current deep learning technology is mainly aimed at light intensity images,and there are few relevant studies on polarimetric images.In view of the above problems,this thesis combines polarization-related theory with convolutional neural networks,and two target classification methods suitable for polarimetric images are proposed.In the absence of high quality polarimetric data,the accuracy of target classification can be further improved.The main research content of the thesis are as follows:(1)The explanation of polarimetric characteristic shows that the polarimetric information distribution is affected by multiple environmental variables and target parameters.We also introduce the principle of convolutional neural network.By virtue of its strong ability of representation learning,it can effectively obtain higher order feature information from a limited number of multidimensional polarimetric images.(2)From the perspective of enhancing the quality of input data,a target classification method based on a polarimetric characteristic factor is proposed.First,the influencing factors of Angle of polarization(AOP)and Degree of linear polarization(DOLP)information are derived from the polarization bidirectional reflection function.On this basis,the information entropy is used to establish the dispersion model of the artificial target and the natural background.Then,the difference in the distribution of polarization information between the two is analyzed.In order to facilitate real-time processing in outdoor applications,the polarimetric characteristic factor proposed in this thesis can calculate the regional dispersion of the target without too many prior parameters,and the pre-processed training images have better classification performance.In addition,in order to reduce the probability of overfitting problems,the method uses a consecutive 2 × 2 convolutions convolutional neural networks(CSC-CNN)architecture that are suitable for shallow neural networks.This architecture reduces the number of weight parameters while improving the non-linearity of the network.After that,two sets of experiments suggest that polarimetric images are more suitable to use dispersion as the feature criterion.Meanwhile,experimental results show that the target classification method based on polarimetric characteristic factor can enhance the accuracy rates by 13.2%compared with the traditional classification method using only ordinary intensity images in different natural backgrounds.In different bands,this method can enhance the accuracy rates by 22.6%.(3)From the perspective of optimizing the architecture of convolutional neural networks,a target classification method based on polarimetric three-dimensional information extraction is proposed.The traditional neural network adopts two-dimensional convolution to extract only two-dimensional information of single-channel images,and the connection between polarimetric multi-parameter images cannot be analyzed.Therefore,the present method takes the polarimetric direction as the information of the third dimension of the polarimetric images,and it analyzes the connection between the polarimetric images with different parameters by the feature extraction of this dimension.The proposed algorithm uses three-dimensional convolution kernels to extract the feature information of the cubic region of three-dimensional polarimetric information,and then completes the subsequent feature extraction through subsequent two-dimensional convolution layers.Finally,experimental results show that the target classification method based on three-dimensional polarimetric radiation information extraction can further enhance the accuracy rates by 6%compared with the traditional classification method that only uses ordinary intensity images in different natural backgrounds.The proposed method can enhance the accuracy rates by 13%in different bands.Meanwhile,it is confirmed that the neural network architecture can occupy as little computing resources and computing time as possible while ensuring that the accuracy rates do not decrease through the comparison of multiple network architectures.(4)In order to verify the recognition of multiple artificial targets in the same natural background,and the better classification results can be achieved by proposed method,the hypothesis was verified by designing experiments.The experimental results show that the multi-target classification method based on the polarimetric characteristic factor can enhance the accuracy rates by 7.3%compared to the traditional classification method using only ordinary intensity images in different natural backgrounds.Meanwhile,experimental results show that the classification method based on three-dimensional polarimetric radiation information extraction can increase the accuracy rates by 15%compared with the traditional classification method using only ordinary intensity images.Therefore,the designed experiments show that the object classification methods proposed in this paper has good generalization ability and high practical application value,and also provides an alternative solution for polarimetric imaging applied to the classification of multiple objects.