Informatica

Multimodal sentiment analysis, due to its comprehensive ability to capture user sentiment, has significant application value in areas such as public opinion analysis. Existing research, however, falls short in several aspects: (1) it inadequately models the global structural information of the image, and (2) it overlooks the potential noise impact within each modality. These limitations hinder the accurate extraction of sentiment cues from individual modalities. To address these issues, we propose a Frequency Domain Decoupling and Semantic Filtering Network for Multimodal Sentiment Analysis. This network primarily integrates frequency-domain decoupling with semantic filtering to process high- and low-frequency image information separately, thereby enhancing model performance. Specifically, we designed a Dynamic Frequency Domain Decoupling Module that applies discrete wavelet transforms for differentiated image processing. This module, combined with a Dual-Domain Loss Function, constrains consistency between the text semantic space and the frequency distribution of the optimized image features, preventing sentiment information loss from excessive filtering. The module also incorporates two key components: a Global Semantic Sentiment Component (GSSC) and a High-Frequency Filtering Component (HFFC). In the GSSC component, we designed a Hybrid Mamba to leverage text in capturing global semantic information from low-frequency image data. Furthermore, our HFFC component generates a dynamic weight matrix guided by text, enabling quantitative noise suppression. Additionally, we developed a Multi-Grained Semantic Purification Module to filter noise at the word, phrase, and sentence levels. Experimental findings from publicly accessible datasets indicate that our proposed model achieves competitive performance compared with existing methods on multimodal sentiment analysis under the adopted experimental settings and sarcasm detection tasks, validating the effectiveness of our noise-suppression method in cross-modal sentiment analysis. A key limitation of the model is its use of DWT: downsampling-related resolution loss in low-frequency subbands and independent subband partitioning compromise the capture of large-scale global structural correlations and local-global feature modelling, which advanced transform techniques can alleviate.

Human Action Recognition (HAR) is an important task in computer vision with diverse applications. However, most existing methods rely on all frames of an action video for classification, which leads to high computational cost and low efficiency. In many cases, a compact set of key keyframes can effectively encode the essence of a complete action. Therefore, this study proposes an efficient HAR method that combines a new keyframe extraction algorithm with a lightweight neural network. Our contribution is three-fold. Firstly, an accurate and efficient key frame algorithm is proposed to alleviate the issue of frame-order confusion in classical clustering methods. Secondly, a key-frame-based multi-feature fusion matrix is constructed to address information loss from spatio-temporal trajectory overlap and the sensitivity issue of viewpoint changes in classical models. Thirdly, a lightweight neural network model is designed to achieve effective convergence within a short training period. The proposed method was evaluated on two public datasets (UTKinect-Action3D and Florence-3D) and a self-collected dataset (HanYue-3D). The experiment results show the advantages of our method in both accuracy and efficiency.