Multimodal biometric systems face critical deployment challenges due to frequent partial modality missing and unreliable modality quality (e.g., blurred fingerprints). Existing methods, relying on unidirectional feature interaction or static fusion, suffer catastrophic performance degradation under such conditions. This paper proposes an adaptive hybrid fusion network (AHFNet) for multimodal hand recognition, which dynamically selects and reweights modalities to correct errors induced by missing or low-quality inputs. The proposed framework comprises three core modules. Initially, a wavelet feature extraction module is implemented to capturing high-frequency biometric feature representations across modalities. Subsequently, a dual-path dynamic fusion module is designed using symmetric cross-attention for bidirectional feature reconstruction, preserving modality specificity via residual projections. Furthermore, this module employs the attention mechanism to perform reweighting, enabling the construction of a dynamic inter-modal compensation pathway at the feature level. Finally, a category-aware dynamic decision fusion module dynamically generates category-level weight coefficients based on input samples, enabling sample-adaptive optimization of multimodal contributions. Experiments on five public hand biometric databases demonstrate AHFNet’s superiority over existing methods across diverse multimodal scenarios, validating its broad applicability. Crucially, AHFNet emphasizes modal reliability differences. Even when certain modalities are missing, it exhibits less than 5% accuracy degradation—a relative improvement of over 75% compared to existing approaches, which typically suffer from more than 20% degradation. This enhanced robustness is consistently observed across all five databases.

Abstract Convolutional neural networks are a hot research topic in finger vein recognition. However, most existing research focuses on increasing the depth and width of the convolutional neural network to improve the network's performance, which has a specific requirement on the computer's computational power. To reduce the computational burden, a lightweight convolutional attention model (LCAModel) is proposed for finger vein recognition to achieve more accurate visual structure capture by exploiting the sensitivity of the attention mechanism to features. First, the attentional model (AModel) is proposed to extract representative features of images, which mainly utilizes the adaptive mapping on space and channels of the convolutional block attention module (CBAM) to make features distinguishable. In addition, considering the integrity of the features, a convolutional model (CModel) is designed to supplement the features of AModel. Finally, the features obtained from AModel and CModel are fused using an adaptive weighting mechanism to make the features more complete. Here, the obtained features are provided into a support vector machine (SVM) for image classification. The experiments are carried out on two publicly available databases, demonstrating that the proposed network structure requires less computing power and performs better.

This paper aims to find a mathematical and statistical way to express natural words' semantic information by mapping words onto a high-dimension continuous space. This paper presents a new approach of training word representations which 1) uses Weighted Local Influence language model (WLI model) by assigning different weight to the words on different locations, 2) introduces polynomial linear regression into the training of word representations, and 3) presents a global optimization scheme introduces polynomial linear regression into global optimal solution word representations. Empirical evidence indicates the availability and effectiveness of the distributed representations based on WLI model. A 7-dimensional sentiment lexicon is constructed based on word representations to prove WLI's semantic descriptiveness and the feasibility of linear empirical functions. Finally this paper discusses over some possible improvements on WLI model in future work.