Research On Image Local Invariant Feature And Its Application

Image feature is the basis for solving many computer vision tasks, andextracting the characteristic parameters that reflect the objective nature is thekey to successfully understand images, correctly interprete and identifytargets. Because local image features are of translation invariance to rotation,scale, viewpoint, illumination and blur changes, they have broad applicationprospects and get a great deal of attraction, such as image stitching, imageregistration, image and video retrieval, object recognition, object tracking anddigital watermarking.The steps for extracting local feature invariance based on visual invarianttheory mainly contain feature detection, feature description and featurematching. The main contributions of this dissertation consists of several parts.(1) It deeply analyses the relevant theoretical basis of the visual invariantfeatures. Through analyzing the mathematics models of the generalizedtransformation group and feature invariance in visual invariant theory, Itexplains the idea of local invariant feature detection and the source of featureinvariance.(2) Through modeling three kinds of typical local image structure,i.e. corner, blob and region, this dissertation specially analyzes the buildprocess from intensity and rotation to scale invariant detector in detail, andsum up a generalized progress of local invariant feature detection.(3)Through modeling several kinds of typical build process of local imagedescriptors, it shows a generalized progress of local invariant featuredescription. It also points out that the descriptors based on Gaussiandifferential operators and feature distribution would have more robust abilityto describe images.(4) There are still some limitations in the original LSSmethod, such that this descriptor is only invariance against small local affineand non-rigid deformations, and insensitive to small translations. To addressthese problems, it proposes several new image descriptors based on LSStexture feature and Cartesian location grid, which are more suitable for different computer vision tasks, such as image matching.(5) To circumventthe presence of many noisy visual words and the hard to define vocabularysize of the traditional Bag-of-Words (BOW) model, this paper concentrateson tuning compact, robust and thus efficient BOW model even with auniversal size for image representation, by using SPLS.Research highlights and corresponding achievements include thefollowing aspects:(1) According to the limitations of the original Local Self-Similarities(LSS,LP) descriptor, it proposes a new framework for extracting texturefeatures, i.e., Local Self-Similarities (LSS,C) and Fast Local Self-Similarities(FLSS,C) based on Cartesian location grid. Different from the natural LSS(LP) descriptor that chooses the maximal correlation value in each bucket toget geometric translations invariance, the proposed LSS (C) and FLSS (C)adopt distribution-based representation to achieve more robust geometrictranslations invariance. In the contexts of image matching and object categoryclassification experiments, the LSS (C) and FLSS (C) both outperform theoriginal LSS (LP), and achieve favorably comparable performance to theSIFT. Furthermore, these descriptors are low computational complexity andsimpler than the SIFT.(2) It proposes a framework to solve instable performance problem, wherethe LSS (C) and FLSS (C) based on distribution-based representation haveinstable performance for photometric and compression transformations.According to the gradient information of differential geometry that has robustphotometric transformations invariance, it proposes two improved LSSdescriptors based on gradient orientation, i.e., Oriented Local Self-Similarities(OLSS,C) and Simplified and Oriented Local Self-Similarities (SOLSS,C).Based on distribution-based representation and differential geometry, theyachieve more geometric and photometric transformations invariance, thus aremore comfortable with the challenges of geometric and photometric changesand get more robust image description capabilities. A large number ofexperiments show that image feature descriptors based on distribution-basedrepresentation and differential geometry will always achieve more robustperformance, and verify the effectiveness of this feature extractionframework.(3) It presents two low-dimensional approaches for extracting distinctive invariant features from interest regions, i.e., PCA and Local Self-Similaritiesfeature based descriptors, namely PCA-LSS and PLSS. They are achieved byapplying PCA on LSS feature field and the improved LSS descriptors ofnormalized patches, respectively. PLSS derived from OLSS (C), thus has asrobust geometric and photometric transformations invariance as OLSS (C). Itdrastically reduces the feature dimension, which makes it more comfortablypractical for limited-memory systems and large-scale computation systems.(4) To circumvent the presence of many noisy visual words and the hardto define vocabulary size of the traditional Bag-of-Words (BOW) model, thispaper concentrates on tuning compact, robust and thus efficient BOW modeleven with a universal size for image representation, by using SPLS. Theproposed approach increases expressive power by employing Sparse PartialLeast Squares (SPLS) for tuning the traditional and high-dimensional BOWmodel and learning more discriminative subspace with10latent variables.Empirical results indicate that the proposed method rebuilds more compactBOW model, yields quite stable results, and outperforms the classical BOWmodels with various vocabulary sizes and PCA with SVM. While thetraditional BOW model appears to be rather sensitive to the variety of thevocabulary sizes. Furthermore, low-dimensional subspace leads to both muchlower computational and memory cost.