Application Of PCA And GA-ELM In Objective Stereoscopic Image Quality Assessment

With the mature of three-dimensional imaging technology, how to evaluate the quality of 3D image effectively becomes one of focus and difficult issues in the field. The method of 3D image quality evaluation generally can be classified into two, one is subjective assessment, the other is objective assessment. The former method scores the 3D image according to the testers' subjective feeling, which is accurate and reliable, but it takes much time and funds and, to some extent, difficult to operate. So establishing an effective stereoscopic image quality objective evaluation standard has become an important research topic in 3D imaging technology field.This paper introduced the background, development status and relative theory of 3D image quality evaluation. For the limitation of human cognition in Human Visual System(HVS), the extreme learning machine(ELM) algorithm is presented for objective stereoscopic image quality assessment in the paper. ELM works for generalized single-hidden layer feedforward neural networks(SLFNs) which randomly choose the input weights and analytically determines the output weights of SLFNs. Compared with traditional neural networks algorithm, ELM not only is easier to select the parameters but also keeps the advantage of extremely fast learning speed and achieves better generalization performance, and is widely applied in the field of function approximations and pattern recognition. For the informative and high complexity of 3D image, we used Principal Component Analysis(PCA) to reduce dimensionality. In addition to, we got the optimized parameters of ELM by genetic algorithm(GA), the so called GA-ELM, which can achieve better performance.In this paper, 362 3D images with different grades are selected, including 121 images as training samples and 241 images as testing samples. Experimental results show that the correct classification rate of 241 different levels of test samples with sigmoid activation function is 93.85%, while the recognition ratio of GA-ELM algorithm can reach 95.85%. Meanwhile, the paper not only studied the effects of different hidden layer nodes for ELM in different activation functions, but also performed an analysis and comparison of the performance of ELM, traditional BP and SVM in stereoscopic image quality assessment. Finally, the performance of ELM and GA-ELM is analyzed and compared simulatively.