Sharing videos on the Internet is becoming increasingly common, therefore, video quality assessment is becoming increasingly important. In existing video quality assessment research, convolution is widely used for feature extraction in the spatial and temporal domains of video, but due to its inherent local perceptual characteristics, the global information perception for video is neglected. At the same time, existing methods do not fully integrate spatial and temporal features. To solve the above problems, this paper proposes a no-reference video quality assessment method that combines convolution with a self-attention mechanism and spatiotemporal fusion. To extract more effective spatial and temporal features, the method first designs a spatial feature extraction module and a temporal feature extraction module, where the spatial feature extraction module enhances the global information of the local features extracted by convolution through long-range dependencies by the proposed feature perception and enhancement module that combines convolution and self-attention. The temporal feature extraction module jointly learns the temporal features of video using Video Swin Transformer and 3D convolution based on video frame sequences and video frame difference sequences. Then, the spatiotemporal feature fusion and enhancement module is used to fuse and enhance the temporal and spatial features through the self-attention mechanism, and finally predict the video quality. Experimental results on the KoNVid-1k and YouTube-UGC datasets show that the proposed method has better performance compared to existing video quality assessment methods, and ablation experiments also verify the effectiveness of each proposed module.

To address the issue of existing deep learning-based segmentation methods for dental models, which overly rely on coordinate features and overlook the differences and complementary nature of geometric and coordinate features, a segmentation network based on attention mechanism and neighbor enhancement is proposed. First, an attention-enhanced module is employed to enhance the neighborhood of input features through different encoding schemes. It designs a differential attention mechanism and a self-attention mechanism to extract comprehensive contextual features of the mesh by considering the properties of geometric and coordinate features separately. Next, a complementary pooling function is constructed to fuse maximum pooling, attention pooling, and average pooling, which promotes the complementary interaction among different features and enhances the robustness of the segmentation network. Finally, an adaptive attention module is embedded to balance the numerical differences among multi-scale features for effective mesh segmentation. Compared with other methods in 200 real patients dental model dataset, the segmentation accuracy of dental model is 96.18%. The experimental results show that the proposed method is superior to the compared deep learning networks and has better performance in the segmentation of dental models.

Existing few-shot Chinese font generation models struggle to effectively decouple font style and content using a limited number of reference characters, leading to suboptimal font generation results. To address this issue, a novel few-shot Chinese font generation method based on contrastive learning and square-block transformation data augmentation is proposed. Utilizing the concept of contrastive learning, the output images undergo separat e style and content comparisons, achieving better adaptive decoupling of font style and content. To further improve the performance of few-shot font generation, a data augmentation method based on square-block transformations is introduced, embedding the global structural information of the fonts through geometric transformations of the square-block areas, thereby enhancing the ability to decouple font style and content. The effectiveness of the proposed method is validated through extensive experiments on public Chinese font datasets. Experimental results demonstrate that, compared to current state-of-the-art few-shot font generation models, the proposed method significantly improves the quality of generated fonts and outperforms comparative models in performance metrics such as FID, SSIM, PSNR, and LPIPS. Notably, the proposed method can be integrated into existing few-shot font generation models and, with only a small training dataset, can significantly enhance the performance of these models.

Interpretable visual image representation learning to reveal image variation factors is a hot research topic in computer vision. Many existing disentanglement methods discover variation factors of images and learn disentangled representations by using extra regularization term. However, it usually leads to an imbalance between disentanglement and generative quality, which affects visual image understanding. To address this issue, a generative visual image understanding method based on disentangled representation learning is proposed in terms of interpretable variations in images. Firstly, a pre-trained Glow model is designed to acquire the latent representations of target images. Secondly, a learning strategy based on image variation is constructed from the latent representations to obtain interpretable directions of candidate traversals. Finally, the contrast module is designed under the contrastive learning perspective to simulate image variations based on the interpretable directions of candidate traversals and then extract disentangled representations. The experimental results show that better results are achieved on the popular disentanglement datasets, which are Shapes3D, MPI3D, Anime, MNIST and Cars3D, where the MIG, DCI, FactorVAE score and β-VAE score metrics reach 0.16, 0.27, 0.89 and 0.98, respectively, on the Cars3D dataset, verifying the effectiveness and feasibility of the proposed method.

To solve the problems of blurring and low resolution in depth maps captured by low-cost depth cameras, a hybrid multimode attention fusion correction network is proposed by combining high-resolution color images in the same scene. To effectively leverage the structure information of color images, the attention mechanism is employed to iteratively align and integrate depth features with color structural features, and the fused features are corrected through the incorporation of color structure features and edge aware maps. Initially, the deep edge aware block combines low-resolution depth maps with color edge maps to extract high-resolution depth edge maps. Then, the side-window edge preserve block is utilized to extract useful structural information from the color image, reducing the impact of extraneous texture information. Finally, multimodal attention fusion correction block utilizes attention mechanisms to extract the structural attention distribution of both depth features and color structural features, aligning and fusing these modalities. The spatial attention weights of the color structural features and high-resolution depth edge maps are used to enhance the supervision of the depth edges, reducing blurring and artifacts. Experimental results on the NYU v2 dataset and Middlebury dataset (2005) show that the average root mean square error test results are 2.96 and 1.45, respectively, which are reduced by 0.10 and 0.08 compared to suboptimal methods. The network can reconstruct high-resolution depth maps with clearer edges and fewer artifacts.