Research On Registration And Classification Algorithms Of Polarization Image

The polarization image can reflect the polarization state space distribution and radiance information of the target,and the change of polarization state reflects the change of the characteristic of the target.This will help improve target detection and recognition.Therefore,polarization images in military,agricultural and other fields have high application prospects.In the aspect of target recognition,it is an important process to classify different objects in images.How to classify the target in polarization image is a research direction.Before the classification of the target,the polarization image is pretreated by image registration at first:For the initial several polarization images,the feature matching is carried out from the view of point feature matching,and the obtained image is then processed in the subsequent processing step by eliminating the mismatch points and combining the random sampling consistency algorithm.Then,encode each pixel of a polarized image into multichannel message,we improved the classification effect from the perspective of classical machine learning and deep learning classification.For the classification of polarized images,the specific work is as follows:The first part is the experiment and improvement of classification from the perspective of machine learning method.On the basis of the classical support vector machine method,the use of spatial-information is added to improve the accuracy of the classification and processing of polarized images.During pretreatment,using three dimensional discrete wavelet transform for image polarization feature extraction,and then through the probability support vector machine(SVM)for each type of probability output,finally,in order to improve classification performance,experiments using the partial correlation between neighboring pixels.The experimental results show that compared with the traditional SVM method,the method proposed in this paper has significantly improved the classification effect.In the second part,the method of deep learning is used to classify experiments.Based on the classification method of 2D-Convolutional Neural Network(2D-CNN),firstly,the local correlation of the adjacent pixels of Markov Random Field(MRF)is combined in the' method,and the experimental results show that the classification effect of polarized images is improved.Next,we combine the three CNN methods of 1D,2D and 3D.The Convolutional Neural Network of 1D,2D and 3D is used to convolve the image block respectively.The output image blocks are then connected by the full connection layer,output layer,and MRF to carry out subsequent classification.Through experimental verification,the improved method of combining 1D,2D and 3D CNN is better than the original 2D CNN for the classification of polarized images.Moreover,the combined methods of 1D,2D,3D CNN and MRF have obtained the best classification results in all the experimental methods.