Research And Application Of Image Matching Based On Label Distribution

With the development of computer and Internet technologies,more and more image information needs to be processed in life and scientific research,such as the Image retrieval,Face recognition and Natural scene detection.It has become the current focus to develop a fast and effective method for finding image information,in order to extract the features of the image and improve the matching accuracy of the image.In addition,label learning related research has also made great progress in recent years,and it has been widely used in image recognition.Defining the information of the image by labeling the image,then select the appropriate method to retrieve,identify and classify the image.Based on this,focusing on the following areas of research.The main research in label learning is the knowledge of label distribution learning.As a new learning paradigm,label distributed learning to reflect the sample of related marks on the importance of the entire sample,as opposed to traditional distributed learning more widely applicable.Due to the small number of sample features in the current label distribution dataset,some label learning algorithms are not highly accurate.Aiming at this problem,a label distribution learning algorithm based on Gradient Boost Decision Tree(GBDT-LDL)is proposed.The main process of the algorithm is to use GBDT to learn and transform the sample features,normalize the new features obtained from the transformation and the original features,and establish the GBDT-LDL model to predict the unknown label distribution.In order to test the prediction ability of the algorithm,six evaluation indicators of marker distribution learning are used to compare the existing classical label distribution learning algorithm.The experimental results show that the GBDT-LDL algorithm is more accurate.Furthermore,the method of statistical hypothesis testing is used to further demonstrate that the algorithm has a good predictive effect.In terms of image matching with HOG to extract image features,this paper introduces in detail the steps and advantages of HOG to extract image information,and proposes two methods to improve the matching effect.One method uses GBDT to construct a classifier to predict image training,and the other method uses the label distribution to apply the cosine similarity measure to improve prediction accuracy.The advantages of the two methods are verified by experimental comparison of decision trees,support vector machines and K-nearest neighbors.The experimental results reflect that the best method is to improve the cosine similarity by the label distribution.Therefore,the experimental scheme of the method of improving the cosine similarity is mainly described and demonstrated in detail.Using ORL face data set to apply image matching experiments and demonstrate the practical application effect of the method.Then performing a detailed analysis of the experimental results as shown in the figure.It shows that label distribution learning can be effectively used in image matching.