Joint Object Extraction Based On Representative Blocks

In recent years, object extraction becomes an important research direction in computer vision field, the object detection and object recognition and other operating methods are endless in order to estimate objects or parts of an image accurately. Nowadays many manipulation methods are generally aimed at one or more objects in an image, and sometimes we find that several related images may contain the same object or portion. Then when given a group of images with relative topics, the common object extraction methods are difficult to realize the estimation of the same object in multiple images. Therefore, the extraction technology of co-objectness emerges as the times require, but there is little research on this aspect.In this paper, a new method of co-object extraction based on the discriminative block is proposed, which is to judge the same parts from two related images. The so-called object is a significant area in the image, that is, the main part of visual perception. Saliency features of the object can be described from five aspects: color, texture, shape, spatial relations and statistical characteristics, while the image effects extracting exist obvious limitations and differences in time complex degree and spatial resolution according to the single feature, so only using one kind of image feature is difficult to meet the actual needs. In order to obtain a clear and accurate object extraction effect, the advantages and the complementary of different image features can be comprehensively utilized.The main research of this article is for the two images given, first, we use Naive Bayesian method to train a Bias classifier, and meanwhile use multi-scale saliency, color contrast, edge density, super pixels straddling and other features to extract a lot of windows that containing it, and then train discriminative patches by linear SVM and related clustering approaches. In particular, conducting an iterative procedure, which alternates between clustering and discriminative training classifier until it converges, and we can get some series of block clusters after many times of iterations. Afterwards we locate the spatial position of the object, and use an image matching algorithm to match similarity of each pair of discriminative patches, and thus select several of the high weighted patches. Finally, we do weight superposition for patches that intersect each object window or contain in it according to the spatial location relationship between the block and the object window, and sort the windows by the score of weights that retained after repeated rejection or merger, at this moment, we get a set of windows that valued maximum weights with accumulation, which contain precisely the similar objects of two images. The experimental result shows that this approach can realize co-objectness estimation effectively.