Research On Quality Assessment Technologies For 3D Mesh Model And Networked Speech

With the rapid development of information technology, the digital multimedia services, such as speech, image, video and 3D mesh model, have been widely used in daily life. User’s subjective experience on the multimedia services mainly depends on the multimedia quality. Therefore, the research on multimedia quality assessment technology is becoming increasingly important. As important multimedia applications, the 3D mesh model and speech offer immersive audio-visual experience to users, and their quality assessment technologies have drawn more and more attention.The quality assessment of 3D mesh model plays a significant role in comparing and optimizing the acquiring and processing algorithms of 3D mesh models. In this dissertation, the objective methods for quality assessment of 3D mesh models are studied, and a full-reference mesh quality assessment method, a no-reference mesh quality assessment method and a mesh denoising algorithm are proposed. Moreover, In order to meet the needs of real-time quality monitoring for the networked speech, two packet-layer models for speech quality assessment are proposed to implement the real-time quality assessment for networked speech in this dissertation, based on the characteristics of networked speech.The major contributions of this dissertation are summarized as follows:1. A quality assessment method of 3D mesh model based on curvature analysis is proposed to accurately measure the distortion of 3D mesh model. The curvature can well describe the visual characteristics of 3D mesh. The proposed method constructs a curvature matrix in the neighbourhood of each vertex, and the local distortion at each vertex is measured by comparing the differences between the singular values of the curvature matrix in both the distorted mesh and the original mesh. The global distortion is further obtained by a weighted combination of the local distortions. Experimental results reveal that the proposed metric not only achieves superior performance in prediction accuracy over all the other competing metrics, but also has very good robustness and stability.2. Considering the fact that the acquired data of 3D mesh model may contain noises, a mesh denoising method and a no-reference mesh quality assessment method based on feature detection are proposed, respectively. On the basis of the bilateral filter, the proposed mesh denoising method modifies the measurement of the spatial distance between two triangles, and adaptively optimizes the denoising parameters of bilateral filter according to the local feature strength estimated by volume integral invariant. Experimental results reveal that the proposed mesh denoising method achieves better performance in preserving sharp features during denoising compared to the bilateral filter. The proposed no-reference mesh quality assessment method utilizes the feature detection method based on the tensor voting theory to perform the feature classification on 3D mesh models, and then estimates the noise level by comparing the differences between the non-edge regions in 3D mesh before and after smooth filtering. Finally, the mesh quality is evaluated according to the predicted noise level. Experimental results show that the proposed no-reference mesh quality assessment method can predict the quality of the noisy 3D mesh accurately.3. Regarding the real-time requirements of the quality assessment for networked speech, a packet-layer model for speech quality assessment, considering the content of the lost packets and the packet loss distribution, is proposed. The proposed model estimates the content of each packet by analyzing the information provided by packet headers, and then extracts the voice packets considerably dominating the speech quality to obtain the coding parameters and parameters related to packet loss of voice segments. Based on these analyses, the coding distortion is firstly estimated using the coding bitrate of the voice segments, and then an equivalent mapping way between random packet loss of voice segments and burst packet loss of voice segments is established to evaluate the distortion caused by packet loss. Finally, the speech quality is jointly determined by the coding distortion and the distortion caused by packet loss. Experimental results demonstrate that the proposed model achieves superior performance over the E model proposed by the international standard G.107.4. A packet-layer model for speech quality assessment based on two-level temporal pooling scheme is proposed to accurately measure the impact of packet loss on the speech quality. The proposed model estimates the frame quality according to the damage level of each frame, and then the speech sequence is divided into speech segments with a variable length, where a short-term temporal pooling is performed to obtain the quality of each speech segment. Finally, the speech quality is determined by a long-term temporal pooling using the quality of each speech segment. More specifically, the proposed two-level temporal pooling scheme predominately emphasizes the impact of the segments with strong impairments on the speech quality, which is in accordance with human auditory perception, and can better reflect the effect of different packet loss distributions on the speech quality. Experimental results demonstrate that the proposed model can predict the speech quality more accurately compared to the existing competing models.