To improve performance of current Non-Reference Image Quality Assessment (NR-IQA) methods, a no-reference image quality assessment algorithm with enhanced adversarial learning was proposed under the latest deep Generative Adversarial Network (GAN) technology. In the proposed algorithm, the adversarial learning was strengthened by improving the loss function and the structure of the network model, so as to output more reliable simulated "reference images" to simulate human visual comparison process as the Full-Reference Image Quality Assessment (FR-IQA) method does. First, the distorted image and undistorted original image were input to train the network model based on the enhanced adversarial learning. Then, a simulated image of the image to be tested was output from the trained model, and the deep convolution features of the reference image were extracted. Finally, the deep convolution features of reference image and the distorted image to be tested were merged and input into the trained quality assessment regression network, and the assessment score of the image was output. Datasets LIVE, TID2008 and TID2013 were used to perform the experiments. Experimental results show that the overall subjective performance on image quality assessment of the proposed algorithm is superior to those of the existing mainstream algorithms and is consistent with the performance of the human subjective assessment.

In order to improve the robustness of Non-negative Matrix Factorization (NMF) algorithm for speech enhancement in different background noises, a speech enhancement algorithm based on Sparsity-regularized Robust NMF (SRNMF) was proposed, which takes into account the noise effect of data processing, and makes sparse constraints on the coefficient matrix to get better speech characteristics of the decomposed data. First, the prior dictionary of the amplitude spectrum of speech and noise were learned and the joint dictionary matrix of speech and noise were constructed. Then, the SRNMF algorithm was used to update the coefficient matrix of the amplitude spectrum with noise in the joint dictionary matrix. Finally, the original pure speech was reconstructed, and enhanced. The speech enhancement performance of the SRNMF algorithm in different environmental noise was analyzed through simulation experiments. Experimental results show that the proposed algorithm can effectively weaken the influence of noise changes on performance under non-stationary environments and low Signal-to-Noise Ratio (SNR) (<0 dB), it not only has about 1-1.5 magnitudes improvement in Source-to-Distortion Ratio (SDR) scores, but also is faster than other algorithms, which makes the NMF-based speech enhancement algorithm more practical.

To improve the universality of image feature extraction, an image feature extraction method based on improved Nonnegative Matrix Factorization (NMF) was proposed. Firstly, considering the practical significance of extracted image features, NMF model was used to reduce the dimension of image feature vector. Secondly, in order to represent the image by a small number of coefficients, a sparse constraint was added to the NMF model as one of the regular terms. Then, to make the optimized feature have better inter-class differentiation, the clustering property constraint would be another regular term of the NMF model. Finally, through optimizing the model by using gradient descent method, the best feature basis vector and image feature vector could be acquired. The experimental results show that for three image databases, the acquired features extracted by the improved NMF model are more conducive to correct image classification or identification, and the False Accept Rate (FAR) and False Reject Rate (FRR) are reduced to 0.021 and 0.025 respectively.

Focusing on the issues that the sparseness of the data obtained after Non-negative Matrix Factorization (NMF) is reduced and the computing scale increases rapidly with the increasing of training samples, an incremental learning algorithm based on graph regularized non-negative matrix factorization with sparseness constraints was proposed. It not only considered the geometric structure in the data representation, but also introduced sparseness constraints to coefficient matrix and combined them with incremental learning. Using the results of previous factorization involved in iterative computation with sparseness constraints and graph regularization, the cost of the computation was reduced and the sparseness of data after factorization was highly improved. Experiments on both ORL and PIE face recognition databases demonstrate the effectiveness of the proposed method.