Research On Ensemble Learning And Its Application In Image Quality Assessment

Ensemble learning,as a research hotspot of machine learning,can significantly improve the generalization performance of the learner under many scenarios.The basic theories,algorithms,and applications of ensemble learning have increasingly become the focus of many machine learning scholars.This dissertation deeply researches and analyzes the problems of data disturbance in ensemble learning,the generation of individual learner and the combination of forecast results.In order to present the research results better,we chose to use image quality assessment as a carrier to comprehensively analyze and implement the application of ensemble learning algorithms in image quality assessment.To sum up,the main content of this paper is as follows:(1)Combined with AdaBoost idea and Bagging idea,an ensemble random forest algorithm was proposed.Ensemble random forest uses random forests as the basic learner under the Ada Boost framework.They can combine Ada Boost ideas and Bagging ideas in the same algorithm,which not only ensures the differences between the learners but also guarantees the accuracy o f the integrated algorithm to predict the quality scores.Using the perceptual features of the image as experimental data,the image quality scores in the LIVEMD image library and the MDID2013 image library are predicted.Experimental results show that the algorithm can not only improve the individual learner's degree of difference,but also enhance the accuracy and generalization ability of the image quality assessment model.Compared with other image quality assessment algorithms,it has certain advantages and is closer to the human experience.(2)Based on the ensemble of random forest algorithm,combined with MultiBoost framework idea and rotating forest(Rotation Forest)perturbation idea,an ensemble random forest image quality evaluation algorithm bas ed on double data disturbance is proposed.Considering the increase of experimental data,how to select valid data in a given data and give full play to the learning ability of the learner becomes a new challenge.The algorithm utilizes rotating forest to perform data perturbation on specific training samples to generate a base learner with a degree of difference,which ensures the improvement of the generalization performance of the ensemble learner.Finally,the learners with certain degree of difference are combined into a complete integrated random forest image quality evaluation algorithm based on dual data disturbances through the MultiBoost framework.In order to prove the effectiveness of the algorithm,experimental data are selected from the multi-scale perceptual features of images in LIVEMD image database and MDID2013 image database.Analysis of experimental results shows that the integrated random forest algorithm based on double data perturbation has certain advantages in image quality evaluation.(3)In order to prove that the ensemble learning algorithm presented in this paper has universal applicability in image quality assessment,this paper proposes a quaternion wavelet transform method for stereo image quality assessment based on ensemble learning.On the 3D image database,the image features are first extracted by quaternion wavelet transform,and the image features are used as the input of the ensemble learning algorithm to obtain the image quality score.Analysis of experimental data can find that ensemble learning also has excellent performance in the evaluation of stereoscopic image quality,and ensemble learning can have a strong generalization in image quality assessment.At the same time,the ensemble learning algorithm proposed in this paper can solve some problems that not easily be solved by other single learners,such as the increasing size of data,which is mainly due to the ensemble learning has greatly improved the generalization performance and scalability of the learner.