Research On Pixel-level Automatic Image Annotation Of Non-rigid Objects Of Interest

Image annotation is a fundamental work in the fields of image processing,computer vision and machine learning,and the quality of annotation directly affects the subsequent model training and algorithm performance.In recent years,with the rapid progress in data acquisition and various requirements of applications,high-precision image annotation,such as pixel-level automatic annotation,has paid more attention and has been the most potential research branch.In view of forestry intelligence,one is usually required to identify non-rigid objects,such as forest "fire flame","smoke",and "cloud",in remote sensing images.Due to the uncertain(or gradient)color and no fixed shape of such objects,the existing annotation methods,such as Bounding Box or image segmentation,perform poorly or even fail to annotate.In this thesis,originating from the independently identical distribution of one-class pixel-pattern objects in machine learning,we propose a pixel-level automatic annotation method for any shaped ROI objects besides the aforementioned non-rigid ones.We highlight our main contributions as follows:(1).A fast extraction algorithm for pixels in convex regions is proposed.The algorithm uses Kronecker product to calculate the positional relationship between image pixels and convex polygon ROI,which can not only automatically realize position determination and pixel extraction,but also has a time complexity of(())O l(9)max n,l,where n is the number of pixels to be determined.l is the number of vertices in the convex region.When n(29)(29)l,the method can be equivalent to as linear method.(2).Aiming at the non-convexity of the image target shape,a fast pixel extraction algorithm for ROI of any shape is proposed.In order to improve the speed of pixel extraction,the non-convex connected region of any shape is decomposed into the union of several disjoint convex sets,and then the method(1)is used to extract pixels.In the decomposition process,two convex decomposition methods are proposed in this thesis,which are called inner decomposition and outer decomposition,respectively.The existence of convex decomposition and the maximum coverage of decomposition are theoretically proved.(3).Based on the pixel samples obtained in work(1)or(2),a fast one-class classification method named L1-SVDD is proposed.The method incorporates the idea of SVDD(Support Vector Data Description)and One-Class classification,while uses the L1 norm to measure the interval.The problem can be solved by linear programming instead of quadratic programming of the original one-class method.Compared with the methods in Bounding Box,Convex Hull,Image Matting and Image Segmentation,the method proposed in this thesis is easy to operate and in line with human vision,the solution speed is fast,and it can achieve non-rigid targets about pixel-level automatic annotation and other features.