A Printed Number SIFT-based Features For Object Recognition

With the real-time monitoring system widely applied in the military field, theindustry and the daily life, the object recognition has become an important technologyin the computer vision field. The core idea of the object recognition is to identify targetby distinguishing and unique features. But on some occasions, owe to the image noise,scale variant and low contrast, it is hard to distingsh features of the target， leading tothe bad performance on the object recognition. Therefore, applied to a printed numbersift-based features for object recognition, which effective solving the problem ondistinguishing target difficultly and improving the performance of the object recognition.In order to achieve the object recognition based on printed numbers, this paper researchon the following areas:Firstly, comparison and analysis on the existed printed number recognitionalgorithm have been done in this paper. The existed printed numbers recognitionalgorithms have bad performance on recognition accuracy and efficiency, especially thatthe accuracy would dramatically decrease when numbers have been deformed. Thispaper applied and introduced a Scale-Invariant Feature Transform algorithm as well asits details and process. Dense SIFT is a dense SIFT algorithm, which improves thequantity and quality of useful information and provides the outlier information.Secondly, in order to improve the recognition accuracy and efficiency of theprinted number recognition, a SIFT algorithm based on the edge detection have beenproposed.This algorithm applies Canny edge detection algorithm to images, whichdetects a collection of keypoints in a short time. After applied SIFT algorithm to thekeypoint collection, it not only shortens the time for searching keypoints, but alsomakes sure that the keypoint describe the image correctly. In this way, we can achievean effective and highly accurate printed numbers recognition system.Thirdly, the model for the printed number recognition has been built. This modelbuilds on the visual words collection based on the clustering result of SIFT vectors,describing images by a bag of visual words. So it can produce a spatial histogram which contains local features. At last, the model classifies the spatial histogram by SVMalgorithm, and the classification is the result of printed numbers recognition.Finally, this paper applied the SIFT algorithm, the Dense SIFT algorithm and theSIFT algorithm based on the edge detection to the proposed model, comparing andanalyzing the results of printed numbers recognition. In conclude, the SIFT algorithmbased on the edge detection will have better recognition accuracy and efficiency.