Image Classification Based On Enhanced Region Detector

Recently, with the rapid development of technology, computers generally haveabilities of high-speed computing and mass-storage, which prompted widespread useof digital image possible. Nowadays, Internet is so popular that the number of digitalimage has increased explosively. In order to effectively manage and organize images,computer needs to exactly understanding the content of images, then manage them incategories. Image classification is an integrated topic that across the patternrecognition, machine learning and computer vision, it also can be seen as a kind oftechnology which can distinguish a set of images based on features of them, and itaccelerate the development process of the image application. Image features can bedivided into global features and local features. The global feature is concern about theoverall of image, such as variance and histogram, etc. It can be easily destroyed byexternal conditions. At present, an image usually represented by local features,compared with the global feature, local features have better stability and differenceswhich can effectively identify the foreground and background images, and they canalso have a better representation of an image by effectively using the feature pointsand information of neighborhood. The image feature extraction is the basis for thework of image classification, which is an important part in the overall of imageclassification, what kind of method you select will directly affect the result of imageclassification.This paper mainly research on the method of local feature detection, aiming toenhance the ability of image representation using local feature and make higher imageclassification accuracy, focuses on Harris, CSS, SIFT, MSER, Harris-Laplace,Hessian-Laplace feature detector, and the advantages and disadvantages of thesemethods are analyzed, then local features are described as feature vectors combiningBOW (Bag of Words) model for image classification. The main contents are asfollows1. Conduct a comprehensive review for the current research of the imageclassification based on image content, make a brief description of research status andthen look forward the application of image classification based on content, introduceseveral methods to detect local features of an image, analyze the exist shortcomings ofthe various methods, laid the foundation for the next phase of Research.2. The paper introducing an image edge information in the detection process tosolve the problem that MSER is sensitive to image edge. Because of the good stability of affine invariant Maximally Stable Extremal Regions has, it is widely used in fieldof the feature detection, it can detect a block structure and the region having afree-form on the image, however, MSER is sensitive to noise, easy to bring someunnecessary pixels to the detected region, which will directly affect the stability andlocating accuracy of the detected region, therefore, the paper introducing an imageedge information in the detection process to solve the above problems. We first detectedge features by canny detector, then using dilated edge to erase irregular region afterbinarization. Experiments show that this method can effectively overcome theshortcomings of the original operator MSER existed, and still maintain a good imagerepresentation.3. Detected features required to be described to eigenvectors which can be usedfor classification, what kind of description method you choose will directly affect thefinal classification results. In this paper, SIFT (Scale-Invariant Feature Transform)descriptors which is widespread used is analyzed, the paper illustrate its effect on theresult of image classification and using SIFT to describe the improved MSER.4. In this paper, the improved local region detector is applied to build an imageclassification framework combined with BOW model, based on an idea that abstractan image into a set of multiple visual words to do image classification, In experiments,scene database and action database are used, compared with the traditional methodusing BOW model and do horizontal comparison with several detection methods, theaccuracy of image classification is increased.