Stereoscopic Image Quality Objective Assessment Based On Deep Extreme Learning Machine

With the continuous progress of stereoscopic imaging technology,how to evaluate the quality of stereoscopic image accurately and stably had been one of the key problems in the technical field of stereoscopic imaging.The method of stereoscopic image quality assessment was divided into subjective assessment and objective assessment.The subjective assessment was the process that testers gave the stereoscopic image score by their subjective feelings.Although this method could obtain accurate results of image assessment,subjective experiment was ambitious,and it was easily influenced by the testers' emotion.So it is very important to establish an accurate and stable stereoscopic image quality objective assessment model.This paper introduces stereoscopic image quality assessment theory?research status and development trend,and proposes to use the deep learning feature extraction and extreme learning machine classification in neural network for establishing an objective stereoscopic image quality assessment model.Extreme learning machine has the characteristics of fast training speed and strong generalization ability.This algorithm has advantages over other similar neural networks,but its random input weights and thresholds of the network will lead to the instable prediction accuracy.So we take advantage of the genetic algorithm to optimize the initial parameters of extreme learning machine for improving the network's classification performance.Because the original stereoscopic image is complex,traditional feature extracting methods are always based on prior knowledge and the single hidden layer network's stability is poor,the sparse auto encoder of deep learning will be used to extract original image features.This method expresses features of the input data layer by layer through the deep learning pre-training process.Then the network not only has a higher classification correct rate,but also possesses a better stability.In this paper,we select 400 images for different quality rank,including 150 training samples,and the others are test samples.The experimental results have shown that after optimizing the initial parameters of the network by genetic algorithm,the classification accuracy is 95.93%,which is 2.1% higher than the not-optimized network's accuracy.Then the proposed deep extreme learning machine network is more stable and efficient.For 250 images of different stereoscopic image quality samples,after testing,the accuracy rate has reached 96.11%,and the network stability is also greatly improved.In addition,we also analyzes the influence of the network about hidden nodes,and compares the performance of the proposed algorithm with the extreme learning machine and the support vector machine in the stereoscopic image quality assessment.