Objective Quality Assessment Of Stereoscopic Images Based On Neural Network Algorithms

With the development of stereo imaging technology,accurate and effective assessment of stereo image quality has become one of the hot and difficult issues in the field of stereo technology.Image quality assessment can be divided into two kinds of methods: subjective assessment and objective assessment.Subjective assessment is given by qualified subjects according to their own feelings during the viewing process.This kind of method can accurately reflect the quality of a stereo image,but it is time consuming and is not easy to implement.Therefore,to establish a set of effective objective quality assessment model of stereoscopic image has become a key research topic of the stereo imaging technology.This paper introduced the research background,development status,development trends and other relevant theories of stereo image quality assessment.Considering the limitations of the present researches on human visual system,this paper proposes a stereoscopic image quality assessment system based on orthogonal locality preserving projection and extreme learning machine.Because of the high complexity and large amount of information a stereo image contains,it is necessary to use appropriate feature extraction method to reduce the dimension of the image.The orthogonal local preserving projection method can preserve the structure of different types of images while reducing image dimension,thus can extract the features of stereo images more effectively.And the extreme learning machine network has the characteristics of simple parameter selection and good generalization.Therefore,we use orthogonal locality preserving projection method to reduce the dimension of the stereo image,and then use extreme learning machine,which is optimized by the genetic algorithm,as the classifier,to obtain better performance of quality assessment system.In this paper,380 stereoscopic images with different distortion processing,covering different grades,are selected.The selected images include 154 training samples and 226 test samples.The experimental results show that with ELM classifier,an accuracy of 93.36% can be achieved by using the method of orthogonal locality preserving projection method,while the accuracy of principal component analysis method is 92.03%.Other results show that by using genetic algorithm to optimize the network parameters,the classification accuracy of ELM network can be improved significantly to 96.03%.In addition,the performance of different neural network classifier is analyzed.