The Object Detection And Recognition Algorithm Based On Object Significance Measurement

With the widespread usage of mobile phones,cameras and other photographic equipments,human beings get into a highly digital information society.The efficient organization and efficient retrieval of large scale of images has become a popular research topic,where object detection and recognition are the important methods to solve this problem.The research on human visual system shows that object detection can act independently of object recognition process.The traditional object detection methods generally use rectangular boxes to classify and identify the potential object areas,but the background interference in the rectangle frame obviously affects the accuracy of the object recognition.The algorithms based on region segmentation often find the segmentation results are not well coincided with the real objects,which also affects recognition result.In this paper,it’s proposed to generate a set of irregular object regions by hierarchical segmentation based on object significance.Then,the features were extracted from the object candidate regions and classified by the algorithm of Bag of Words model.First,in traditional edge detection methods,the brightness edge model could not detect the edge between textural regions,pure texture model could not effectively detect the brightness of the edge.This paper used the changes in a number of clues to calculate the score,which included the joint brightness,color,texture.The histogram difference operator was used to compare the differences between the two half-disk to calculate the boundary probabilities of each pixel using the formula.Secondly,due to the structural regularity of man-made environment,geometric context information was calculated based on linear detection and vanishing point estimation.Since the straight line disappears in three orthogonal directions,the straight lines were divided into three significant directions and the vanishing points were estimated by the EM algorithm.Then,the geometric context information was obtained by partial calibration and relative rotation estimation with vanishing point constraints.Again,because many of the object appearances are approximately homogeneous,edges and geometries can often be used to restore significant object occlusion boundaries.Although it is not possible to expect to locate each object,a small candidate regions including most objects can be generated.We used the information such as boundary,geometry,layout,color,texture and so on to guide the hierarchical segmentation,generated a group of irregular potential object candidate regions.The candidate regions were sorted by structured learning so that the regions with the highest ranking were most likely to correspond to different objects.Finally,the image feature was extracted by SIFT for the candidate object region after sorting,then used the Bag of Words model to describe the characteristics and generated dictionary.The SVM classifier was used to classify and identify the word packs of each image as feature vectors.On the VisualStudio + OpenCV platform,the object detection algorithm was developed to generate irregular potential object candidates and sorted them.Then the object recognition algorithm based on the model of Bag of Words was used to classify and get the result.Experiments shew that the performance was promising.