Research On Hyperspectral Image Band Selection Algorithm

Hyperspectral Image records the reflectance information of different scene objects w.r.t.electromagnetic waves in different wavelengths.Benefitted from the above characteristics,hyperspectral image and related processing techniques have been used in a wide range of applications such as medical image analysis and product quality inspection,and achieved great success both in scientific researching and commercial development.However,the high dimensionality and large redundancy of hyperspectral image will lead to immense spatial and temporal complexity especially for algorithms used in traditional optical imagery.Hence,how to extract the discriminative spectral information and reduce the data dimensionality is becoming an important researching topic.Band selection is a common method to reduce the dimensionality,and meanwhile filter the redundant information for hyperspectral image.Compared to feature extraction techniques,band selection only select some representative bands,but will not change the physical property of them.As a result,it is more understandable and more practical in various applications.Nowadays band selection techniques can be roughly categorized into group-wise and point-wise methods.Group-wise methods first separate all of the bands into groups,in order to minimize the intra-group distance while maximize the inter-group distance.Then they select one representative band in each group to generate the band selection result.Group-wise selection can prevent the selected bands from being highly correlated,but since the relation among those bands are not explicitly considered,the result is usually not optimal.Point-wise selection achieves the band selection result via solving a combinatorial optimization problem,by which the subtle relation among bands can be better exploited.But due to the large complexity of combinatorial optimization problem and the instability of the corresponding optimization algorithms,the selected bands are sometime highly correlated.In consideration of the advantages of the above two category of methods,this thesis proposes two algorithms,each belongs to one of them.Among these algorithms,the group-wise one exploits the prior that neighboring bands are usually with stronger relation.It constrains that only bands with contiguous wavelengths can be grouped together,and further proposes a dynamic programming based optimization method to solve the clustering problem.This algorithm can assure the global optimality of the solution and is with lower temporal complexity.The point-wise algorithm designs a linear reconstruction mechanism to capture the subtle relation among various bands,and uses an adaptive noise reduction method to minimize the influence of noisy bands.The experiments demonstrate that the proposed algorithms can achieve a stable performance in various datasets,and hence are more practical in real-world applications.