Medical Large Vision-Language Models (MLVLMs) show encouraging results in medical diagnostics but easily in-herit biases from pretraining data, leading to bias catastrophic inheritance, where data biases persist and distort predictions. In this work, we present the first systematic study of this issue in MLVLMs, revealing how inherited biases affect classification and free-text reasoning tasks. We propose Logit Fairness Adjustment (LFA), a training-free debiasing method that operates at the logits level to recalibrate biased predictions. LFA quantifies bias by computing logit margins between valid and invalid medical images, applying logit smoothing when the margin is small to reduce overconfidence and bias compensation when the margin is large to reinforce valid features. We introduce the Medical Multimodal Bias Benchmark to assess bias severity across binary classification, multi-class classification, and free-text reasoning. Experiments on LLaVA-Med, SkinGPT-4, and Qwen-VL-7B show that LFA effectively mitigates bias for MLVLMs.

Multivariate time series forecasting has a significant role in numerous areas of the real world. Despite the noteworthy accomplishments of Transformer-based models in this domain, there remains a significant challenge in effectively capturing temporal dependencies and multivariate correlations. In previous studies, researchers have employed the entire sequence of each variable as a token and captured multivariate correlations by means of self-attention. However, this approach of embedding the entire sequence projection into variable tokens limits the model’s ability to extract deeper temporal features and learn complex temporal patterns. In this work, we propose GLoSyformer, a Transformer-based model. Specifically, we introduce double sampling embedding to enhance the representation of variable token temporal features from both global and local perspectives, and utilise the TimesBlock module to capture global and local temporal dependencies in parallel. Additionally, we propose a sparse variable attention mechanism that captures multivariate correlations while reducing the interference of irrelevant variables. Finally, experimental results on eight real-world datasets demonstrate that GLoSyformer outperforms existing models and achieves superior forecast performance.

Time series forecasting plays a crucial role in a wide range of applications, including stock market analysis, weather forecasting and power load prediction. Recently, Convolutional Neural Network (CNN)-based and Multilayer Perceptron (MLP)-based models have made significant progress in time series forecasting and have become strong alternatives to Transformer-based models. However, despite their impressive performance, these methods have not yet reached their full potential. In addition, the sequence components derived from existing decomposition techniques often remain highly complex, limiting the performance gains achievable by these models. To address these issues, this paper proposes the Bi-Frequency Modeling Network (BiFMNet), a novel framework that adaptively decomposes time series data into low-frequency and high-frequency components. Using an adaptive frequency separation strategy, BiFMNet tailors the decomposition to the unique characteristics of each channel. In the low-frequency modeling component, we designed an MLP module based on a multi-scale Exponential Moving Average (EMA) to capture long-term trend information within the low-frequency data. In the high-frequency modelling section, we build an efficient high-frequency CNN module designed to learn the periodic fluctuations of high-frequency information. Through experimental evaluation on seven publicly available benchmark datasets, BiFMNet significantly outperforms current mainstream time series forecasting models on several key metrics, demonstrating excellent performance and strong generalisation capabilities.

While view-based methods have achieved significant advancements in the classification of rigid 3D models, they face substantial challenges when applied to non-rigid 3D model classification. The inherent deformations of non-rigid 3D models, such as bending, stretching, and compression, present challenges for traditional methods of generating 2D views, which typically capture only a single type of information, such as lighting or depth. Consequently, these methods struggle to effectively capture the intrinsic features and key characteristics of non-rigid 3D models. To address this limitation, we propose a novel multi-view deep learning framework for the classification of non-rigid 3D models. Unlike existing methods that generate 2D views from limited information, our approach generates 2D views based on geometric spectral features. Specifically, we utilize multiple Laplace-Beltrami eigenfunctions at different frequencies to render a series of 2D views of the model, which are subsequently fused into a unified input space. In addition, we introduce a view aggregation method based on spatial positional relationships. This method selects adjacent views and employs an attention mechanism to compute the correlations between them, thereby facilitating information exchange among neighboring views. This design effectively exploits the latent inter-view information dependencies and enhances the representational capacity of the model. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods, confirming its robustness and effectiveness for non-rigid 3D model classification.

Infrared and visible image fusion is an extensively investigated problem in infrared image processing , aiming to extract useful information from source images. However, the automatic fusion of these images presents a significant challenge due to the large domain difference and ambiguous boundaries. In this article, we propose a novel image fusion approach based on hybrid boundary-aware attention, termed HBANet, which models global dependencies across the image and leverages boundary-wise prior knowledge to supplement local details. Specifically, we design a novel mixed boundary-aware attention module that is capable of leveraging spatial information to the fullest extent and integrating long dependencies across different domains. To preserve the integrity of texture and structural information, we introduced a sophisticated loss function that comprises structure, intensity, and variation losses. Our method has been demonstrated to outperform state-of-the-art methods in terms of both visual and quantitative metrics, in our experiments on public datasets. Furthermore, our approach also exhibits great generalization capability, achieving satisfactory results in CT and MRI image fusion tasks.

In this work, we propose LiteMVGNet, a novel lightweight neural network for 3D shape classification. It is based on depth maps generated by multi-view rendering of the corresponding 3D model. LiteMVGNet is designed to be lightweight and effective in various aspects. First, the views and corresponding depth maps are partitioned into groups. Next, depth map features for each group are separately extracted by an adapted MobileNetV2 block. Finally, the extracted group features are fused by an adapted MobileViT block. The views are partitioned by good geometrical semantics and ECAnet is utilized to facilitate extraction of effective features. As demonstrated by experiments, in comparison with the state-of-the-art benchmark models, the proposed one cuts the network parameter count by a third and more and reduces the floating-point operation count by even one or two orders of magnitude. Still, the proposed model yields classification accuracies comparable with the benchmark models.

Acquiring semantics directly from a point cloud is an important requirement for handling point cloud tasks. However, point clouds captured with laser scanner equipment are often incomplete due to the limitations posed by target occlusion and light reflection. Consequently, recovering the complete point clouds from partial and sparse ones is essential for further studies. In this paper, we model a novel projected generative adversarial network (PGAN) for point cloud completion. First, we present a multi-scale generator module (MSGM) to fully capture the local structures and global shape in the raw incompletion point cloud and generate the multi-scale complete point cloud. In contrast to existing point cloud feature extractors, our MSGM promotes a correlation between different regions of an incomplete point cloud and integrates the contextual information of the point cloud. Second, we observe that the existing point discriminator is inadequate to enhance the discrimination of the prediction point cloud. To address this problem, we project the completed point cloud to 2D maps and apply adversarial training to discriminate the geometrical shape from a specific viewpoint. Comprehensive experiments on the ShapeNet and ModelNet40 datasets show that the proposed method performs well against existing point cloud completion tasks. We also present an ablation study to demonstrate the advantages of the projected generative adversarial network.

The detection of 3D interest points is a central problem in computer graphics, computer vision, and pattern recognition. It is also an important preprocessing step in the analysis of 3D model matching. Although studied for decades, detecting 3D interest points remains a challenge. In this study, a novel multiscale bilateral filtering method is presented to detect 3D interest points. This method first simplifies repeatedly the input 3D mesh to form k multiresolution meshes. For each mesh, on the basis of the computed saliency of the mesh vertex, the bilateral filtering is used to remove the noise of the mesh saliencies and the global contrast to normalise the saliencies, and then the interest points are extracted on the basis of the normalised saliency. The proposed method then gathers and clusters all interest points detected on the k multiresolution meshes, and the centres of these clusters are treated as the final interest points. In this method, both the spatial closeness and the geometric similarities of the mesh vertices are considered during the bilateral filtering process. The experimental results validate the effectiveness of the proposed method to detect 3D interest points. This method is also tested the potential to distinguish 3D models.

Point set registration is a fundamental problem in many domains of computer vision. In previous work on the registration, the point sets are often represented using Gaussian mixture models and the registration process is represented as a form of a probabilistic solution. For non‐rigid point set registration, however, the asymmetric Gaussian (AG) model can capture spatially asymmetric distributions compared with symmetric Gaussian, and the structural feature of the point sets reserve relatively complete and has important significance in registration. In this work, the authors designed a new shape context (SC) descriptor which combines the local and global structures of the point set. Meanwhile, they proposed a non‐rigid point set registration algorithm which formulates a registration process as the mixture probability density estimation of the AG mixture model, and the method introduce the structural feature by the new SC. Extensive experiments show that the proposed algorithm has a clear improvement over the state‐of‐the‐art methods.

Content‐based image retrieval (CBIR) is a research hotspot. To improve the performance of a CBIR system, especially the retrieval accuracy, this work proposes a method that uses a soft hypergraph combined with a weighted adjacent structure (WAS) to retrieve images. In this method, the similarities between images are computed and a similarity matrix is constructed by a conjoined colour difference histogram and micro‐structure descriptor method. Furthermore, a novel WAS and a soft hypergraph model are utilised to further improve the retrieval precision. The proposed method is compared with other methods in several datasets. Experimental results manifest the performance and robustness of this proposed method.

This paper presents a reference sharing mechanism-based self-embedding watermarking scheme. The host image is embedded with watermark bits including the reference data for content recovery and the authentication data for tampering location. The special encoding matrix derived from the generator matrix of selected systematic Maximum Distance Separable (MDS) code is adopted. The reference data is generated by encoding all the representative data of the original image blocks. On the receiver side, the tampered image blocks can be located by the authentication data. The reference data embedded in one image block can be shared by all the image blocks to restore the tampered content. The tampering coincidence problem can be avoided at the extreme. The maximal tampering rate is deduced theoretically. Experimental results show that, as long as the tampering rate is less than the maximal tampering rate, the content recovery is deterministic. The quality of recovered content does not decrease with the maximal tampering rate.

Graphic matching is a very critical issue in all aspects of computer vision. In this paper, a new graphics matching algorithm combining shape contexts and reweighted random walks was proposed. On the basis of the local descriptor, shape contexts, the reweighted random walks algorithm was modified to possess stronger robustness and correctness in the final result. Our main process is to use the descriptor of the shape contexts for the random walk on the iteration, of which purpose is to control the random walk probability matrix. We calculate bias matrix by using descriptors and then in the iteration we use it to enhance random walks’ and random jumps' accuracy, finally we get the one-to-one registration result by discretization of the matrix. The algorithm not only preserves the noise robustness of reweighted random walks but also possesses the rotation, translation, scale invariance of shape contexts. Through extensive experiments, based on real images and random synthetic point sets, and comparisons with other algorithms, it is confirmed that this new method can produce excellent results in graphic matching.

A new method for constructing triangular patches to pass the C1interpolation conditions (boundary curves and cross-boundary slopes), on the boundary of triangles is presented. The triangular patch is constructed by a basic triangular operator and an error triangular operator. The basic operator is a polynomial of degree six, which approximates the interpolation conditions with a higher approximation precision, while the error operator is constructed by the side-vertex method, which passes the C1error boundary conditions. The C1error boundary conditions are formed by the C1interpolation conditions minus the boundary curves and cross-boundary slopes taken from the basic operator. The basic operator and the error operator are put together to form the triangular patch. Comparison results of the new method with other two methods are included.