Distortion Image Invariant Recognition Based On Moment Invariants And Local Binary Pattern

Image distortion is a phenomenon that the geometric structure of images has been changed.This phenomenon is the result of the image scaling,translation and rotation in the process of the imaging.The shape,size and orientation change of objects is inevitably introduced,since the different angle and position by using observation tools to taking images from the real world.It is more difficult for computer to recognize the objects of image.Therefore,distortion images invariant recognition has been widely investigated in the field of image recognition.This thesis mainly studies the distortion image invariant recognition problem in term of image global feature methods invariant moments and image local feature methods Local Binary Pattern(LBP).The proposed methods are listed as follows:1.A direct method to obtain rotation,scaling,translation(RST)invariants from Krawtchouk moments is designed in this thesis.This global feature method is named with Explicit Krawtchouk Moment Invariants(EKMIs).The existing Krawtchouk moment invariants were derived by a linear combination of geometric moment invariants with weak noise robustness.To eliminating this undesirable effects,this thesis develops a mathematical framework for establishing the relationship between the Krawtchouk moment invariants of distortion image and Krawtchouk moments of the original image.The images which are simple background and single object in Coil-20,Coil-100 and Butterfly public database are used in the experiment.Theoretical and experimental results indicate that the proposed method improves distortion image recognition accuracy and has higher noise robustness.2.This thesis proposes a local feature method named two dimensional Local Binary Pattern(2DLBP)for texture images invariant recognition.In conventional LBP descriptor,the extracted feature was shown in the form of histograms result and ignored contextual information between these LBP values.The proposed method firstly takes the advantage of uniform rotation invariant LBP to describe the image so as to obtain the corresponding LBP feature map.Secondly,the weight contextual information in a LBP feature map is extracted throughout introducing sliding window in the fixed size.Finally,the radius parameter of the LBP descriptor is changed to construct the multi-resolution 2DLBP feature of the image.Theoretical validation shows that the proposed algorithm is general and can be integrated with other variants of LBP to derive a new image recognition method.At the same time,compared with the state of the art LBP variants methods,experimental results show that the proposed method has more effective image recognition ability.