The Research On FTIR Microscopic Imaging Classification Algorithms Based On Spatial-spectral Features

FTIR microscopic imaging is a powerful technique that combines the analysis of both spectral and spatial information with high spectral resolution.With a detectability of 100~400 spectral channels and spectral resolution reaching ?/??=100 orders of magnitude of spectral images,this process is known as hyperspectral imagery.Thus,FTIR microscopic imaging is also known as infrared hyperspectral microscopic imaging.Spectral information and qualitative analysis studies always focus on the spectral dimension,because microscopic imaging data have abundant diagnostic information.Nowadays,spectral analysis models develop extremely rapidly.However,due to the spectral chemical specificity of FTIR microscopic images,there are still some serious problems in qualitative analysis of spectral visualization.Traditional FTIR microscopic classification methods have obvious shortcomings,that is,only relying on spectral information in classification,ignoring the positive role of spatial texture information features.Over the last decade,RSI technology has received significant attention for exploiting spatial features to improve the classification performance in RSI imagery.Therefore,inspired by this,the spatial texture information is incorporated into the classification process of FTIR microscopy,which provides supplementary information about the spectral properties and effectively improves the classification performance.Therefore,a spatial-spectral structure that takes into account the existing spatial texture relationships between image pixels within a high-dimensional spectral space is likely to improve the classification accuracy.The main contents are as follows:(1)In view of the lack of physical meaning in the raw spectral features of the multivariate methods,spectral feature extraction is introduced into the subspace partition based on the absorption peak of univariate analysis theory to remove the spectral redundancy band.Combining univariate and multivariate analysis ideas,we propose a CAPI algorithm.This investigation mainly implements four developed strategies for selecting feature bands,which include ASBR,SSR,SP~3 CA,and SP-OIF to extract subsets of the absorption peak bands from different functional group intervals.The CAPI algorithm applies spectral fingerprint characteristics and chemical functional group information of the microscopic image in a more comprehensive manner,and selects representative band subsets from the original FTIR microscopic image data to achieve the objective of dimensionality reduction.The new classification model is constructed using the band subset as the feature matrix,which provides an important basis for optimizing the spatial distribution information of sample components.(2)Aiming at the noise sensitivity of probability measure of partition function,we have presented a new probability measure method based on the NSC to extract texture features.The spatial feature extracted by NSC is first stacked with spectral features of characteristic absorption peak internally,and then this spatial-spectral structure is used to update feature information for LSSVM classification.The comparison results show that the accuracy of this spatial-spectral structure is higher than that of other multivariate analysis methods.The proposed NSC method could effectively reduce the influence of speckle noise and preserve texture information in image classification.Meanwhile,the convergence rate of the weight factor q is not increased,and the range of linear intervals satisfying scaling invariance is large.(3)To solve the problem of increasing scaling variable intervals,two solutions are proposed: 1)Scaling invariant interval pruning strategy is adopted.In practical application,the scale invariance criterion is too harsh.Because of the irregular distribution of texture surface and various pixel noises,the partition function cannot satisfy the self-similarity in the infinite region,and only has scale invariance within a certain range.In order to solve this problem,the scaling interval pruning strategy is adopted to prevent the occurrence of texture feature distortion.2)Two parameter indicators are proposed for the local linear anomaly of the scale invariant interval.The size of template scale ? and the cut-off threshold of weight factor q are set.The linear fluctuation degree of the partition function SELF_Q(?) and the convergence coefficient of the weighting factor WFCC_S (q) are proposed.The two indicators search for the optimal size of scale ? and the optimal convergence value of weight factor q to avoid scale invariant interval anomalies.At the same time,the visual graph of linear interval curve is designed.(4)Considering the shortcomings of traditional spatial model with more parameters and complicated operation,the AVWRF directional operator based on the DFBIR field model is proposed for extracting spatial characteristics of FTIR microscopic imaging.The method relies on DFBIR theory.The AVWRF field is used to describe the irregular,random,and highly complex shapes of natural objects such as coastlines and biological tissues,for which traditional Euclidean geometry cannot be used.Firstly,the AVWRF direct ional operator is used to characterize the complex characteristics of texture details at different levels,and the multi-directional spatial feature information is incorporated into the spectral features.Secondly,probabilistic principal component analysis extracts spectral features,and then the spatial features of the proposed AVW directional operator are combined with the former to construct a spatial–spectral structure.The proposed AVWRF direction operator achieves the classification accuracy while achieving less space for setting parameters and preserving spatial characteristics with less feature dimensions.(5)The FTNIR imaging spectral data of the highly mixed drug tablets is severely aliased,and the spectral chemical specific variation directly causes the identification of many different components of the tablet.The EC2 MAPs algorithm was proposed to extract the spatial geometric structure features of images.With the help of EC2 MAPs,several convex body models are constructed to obtain the spatial feature vectors,and describe the spatial lacunarity roughness information of image structures.Then the spatial information is integrated into the classification strategy of a posterior probability SVM by using MCIKL framework,which provides the possibility for feature fusion of spatial roughness information.The spatial-spectral composite kernel is constructed to achieve more satisfactory experimental results than single kernel,and improve the classification accuracy of single kernel.