Image Local Feature Extraction Based On Orthogonal Moment

Image features show the essential information of an image.Generally,image features,according to whether taking whole complete image into consideration,can be divided into two sub-fields: complete image features and local image features.Normally,local image features are robust to image occlusion,and shows high superiority in image deformation(such as image affine,image illumination,image blur etc).But their image descriptive ability is a little lower since the natural disadvantage of local image descriptiveness.Further,their image recogonition ability will be decrease when suffered from image noise or image blur attack.In contrast to local image features,complete image features possess much stronger descriptive ability,and show much higher robustness against image noise or image blur attack.Sadly,their robustness will be dramaticly decreased when suffered from image occlusion.Moreover,complete image features are powerless to deal with images with complex background.Orthogonal moment is another complete image feature whose excellent qualities are inherited.Besides,Orthogonal moment itself also has many excellent characteristics,such as low computation cost,conveniently sampling and powerful image descriptiveness.This thesis designed an orthogonal moment-based approach for image local feature extraction.Mainly,the method has two parts inside :constructing image feature dectector and constructing image feature descriptor.The former is applied for locating image interest points(also named as key points or feature points).The later,a group of feature vectors,is used to uniquely represent the region of interest points.The feature detector of the proposed approach is constructed by the following steps.Firstly,image hessian matrix is generated by orthogonal moments.And cornerness is defined by the difference of hessian matrix's determinate and trace's square.Secondly,with the theory of image scale space,image pyramid is imitatively constructed by multiscale-layer orthogonal moments to achieve features' scale invariance.Thirdly,in order to get the biggest cornerness response in each pixel's 3x3 neighbor,top neighbor and bottom neighbor,non-maximum suppression is applied to obtain image corner responses under multi-scale,thereby candidate feature interest points are located.Fourthly,2D-parabolic interpolation is employed to locate the interest points precisely at the sub-pixel level.In this step,three coordinates(interest point and its top and bottom neighbor center)are selected to fit a 2-order function,and the peak is considered as accurate location.Finally,with the help of principal curvature,edge responsing points' disturb is eliminated by comparing the radio(difference of trace's square and determinate)and the threshold.The feature descriptor of the proposed approach is generated by the procedures as follows.Firstly,interest points' 4-by-4 neighbors are divided into 16 regions.Secondly,8 directions are generated by calculating gradient histogram,among which the one who scores highest is considered as dominant orientation.Finally,128 dimision feature vectors are constructed by 4-by-4 neighbors' 8 directions of feature interest points.Experimental results illustrate the following conclusions.Firstly,when encountered with rotating and zooming attack,the repeatability of the proposed method shows less volatility,1-precision and recall are tied with other methods.Secondly,when encountered with viewpoint changing,the proposed method has a high repeatability score and shows the tendency of superiority over other methods.Thirdly,when encountered with image blur,the proposed method gets a higher repeatability score,and the repeatability acurve waves less.Moreover,1-precision and recall are higher than the others in general.Fourthly,when encountered with illumination changing,the proposed method scores highest in repeatability and recall,from which shows the superiority over other methods.Fifthly,when encountered with JPEG-compression,the proposed method shows dominance in this aspect.Finally,when encounter with image noise,the proposed method alse gain the highest repeatability score.In summarize,comprehensive experimental results show the method-proposed can effectively resist on image deformation,and be more robust to image quality changes.