The Study On The Saliency Of Region

Image feature region detection is an important research area in computer vision and pattern recognition.It has been studied for several years.Feature region detection,as the basis of the other research areas,decides the final results of the following applications.Because the traditional methods of feature detection are relatively independent of the following applications,the methods are unable to self-adaptively select the effective features for the application according to the training dataset.Inspired by the model of the human visual attention,the dissertation proposes an application-oriented definition of saliency,which is capable to evaluate the effectiveness of the feature for the particular application.The experimental results show that the salient regions achieve better results than those unsalient ones.As a basic research area,the study of the saliency for a region is important for improving the results of the related applications.The main research work and innovations of this dissertation are listed as follows:A matching-oriented algorithm of saliency for a region is proposed.The proposed algorithm takes the matching rate of the regions into account when calculating the saliency.Based on the proposed algorithm,the salient regions achieve better matching rate than unsalient ones.Also,the proposed algorithm gives a way to estimate the probability density functions(pdf)of the multi-dimensional features.Concretely,the dissertation defines the saliency of a region as the Kullback-Leibler(K-L)divergence between the pdf of correct matches to the given region and the pdf of incorrect matches.With the help of the statistics of the natural images,the algorithm introduces elliptically symmetric distribution(ESD)and log-normal distribution(LND)to model the pdf of multi-dimensional features.Then,the calculation of the K-L divergence based saliency is given.The experimental results on several natural image datasets verify the effectiveness and the superiority of the proposed method.Additionally,the results show that the saliency of a region is self-adaptively adjusted when the training set is changed.A calculation of saliency for a region via sample-based K-L divergence estimation is proposed.Due to the previous definitions always assume that the distributions of the features in images follow a few classes of pdfs,these definitions may not be suitable for the new training sets.The proposed method directly uses the samples to calculate the K-L divergence,so that the method is suitable for a much broader range of images.The proposed method has two major advantages: 1)the sample-based K-L divergence estimation dose not have to add restricts on the involved pdfs,thus the proposed algorithm is more flexible,2)the saliency is calculated without an intermediate estimation of the joint pdfs,thus the calculation is simplified.The experiments are conducted on the natural scenes,texture images and artificial images.The results show that the proposed method is able to calculate the saliency of the regions in different classes of images.Furthermore,the results suggest that the proposed method is also valid for different feature descriptors.A recognition-oriented algorithm of saliency for a region is proposed.The algorithm takes the representability,the discrimination and the matching rate of a region into account.The salient features can be used in matching-based object recognition and get better recognition rate.The proposed algorithm defines the representative regions as those which exist in the interest object.The discriminative regions are the representative regions which are away from the background.Using the sample-based K-L divergence estimation algorithm,the discrimination of a region is calculated.With the help of the matching-oriented definition of saliency,the proposed algorithm brings the matching rate into the estimation of the saliency.Thus,the salient regions have both high discrimination and matching rate.The experimental results show that the salient regions,detected by the proposed algorithm,are useful for discriminating the object from the background,but also improve the accuracy of object recognition.