Yutong Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yutong Li has a total of 100 co-authors. Here are the top co-authors by number of shared publications:
- Zhengjun Liu
Harbin Institute of Technology
52 shared publications - Shutian Liu
School of Physics Harbin Institute of Technology Harbin China
26 shared publications - Bin Gao
School of Computer and Big Data Heilongjiang University Harbin China
17 shared publications Shutian Liu
17 shared publications- Ziyang Li
School of Physics, Harbin Institute of Technology, Harbin, China
16 shared publications - Guancheng Huang
School of Physics Harbin Institute of Technology Harbin China
14 shared publications
Latest Publications
- 17 November 2026
- Published by Wiley in Laser & Photonics Reviews
High‐Fidelity, Adaptive Label‐Free Imaging via Color‐Multiplexing With Analytical Aberration Estimation
Modern biomedical imaging increasingly demands efficient and accurate reconstruction of multidimensional optical fields to address complex tissue‐level questions. However, conventional image‐domain optimization methods often face trade‐offs among aberration correction, data redundancy, and reconstruction fidelity. In addition, modeling nonlinear optical systems using linear time‐invariant assumptions introduces mismatch that limits the effectiveness of post‐processing, such as vignetting correction. This work presents color‐multiplexed adaptive phase imaging (CAPI), a computational imaging framework that combines high‐throughput data acquisition, analytical aberration estimation, and Kramers–Kronig‐guided gradient regularization. CAPI estimates wavelength‐dependent aberrations by leveraging phase inconsistencies in spectrally overlapping regions, using edge‐aware segmentation and Zernike polynomial fitting to enable non‐iterative estimation of multi‐channel optical distortions. By expressing the reconstruction cost using first‐order spatial gradients, CAPI enforces amplitude‐phase consistency and mitigates artifacts caused by model mismatch. Only five raw RGB images are required to achieve high‐resolution imaging with aberration correction across three spectrum channels, CAPI reduces the acquisition burden to 6.7 of conventional methods while maintaining high reconstruction accuracy. These innovations provide a scalable and efficient platform for high‐throughput biomedical imaging, particularly in resource‐constrained or data‐sensitive scenarios.
- 17 November 2026
- Published by Wiley in Laser & Photonics Reviews
High‐Fidelity, Adaptive Label‐Free Imaging via Color‐Multiplexing With Analytical Aberration Estimation
Modern biomedical imaging increasingly demands efficient and accurate reconstruction of multidimensional optical fields to address complex tissue‐level questions. However, conventional image‐domain optimization methods often face trade‐offs among aberration correction, data redundancy, and reconstruction fidelity. In addition, modeling nonlinear optical systems using linear time‐invariant assumptions introduces mismatch that limits the effectiveness of post‐processing, such as vignetting correction. This work presents color‐multiplexed adaptive phase imaging (CAPI), a computational imaging framework that combines high‐throughput data acquisition, analytical aberration estimation, and Kramers–Kronig‐guided gradient regularization. CAPI estimates wavelength‐dependent aberrations by leveraging phase inconsistencies in spectrally overlapping regions, using edge‐aware segmentation and Zernike polynomial fitting to enable non‐iterative estimation of multi‐channel optical distortions. By expressing the reconstruction cost using first‐order spatial gradients, CAPI enforces amplitude‐phase consistency and mitigates artifacts caused by model mismatch. Only five raw RGB images are required to achieve high‐resolution imaging with aberration correction across three spectrum channels, CAPI reduces the acquisition burden to 6.7 of conventional methods while maintaining high reconstruction accuracy. These innovations provide a scalable and efficient platform for high‐throughput biomedical imaging, particularly in resource‐constrained or data‐sensitive scenarios.
- 3 November 2026
- Published by Optica Publishing Group in Optics Letters
- Vol. 51 (22) , 6179
High-quality lensless on-chip microscopy via mixed coherence control with H&E-like pseudocoloring
Achieving high-quality lensless microscopy remains challenging because of an inherent coherence trade-off: lasers enhance resolution but introduce artifacts, while LEDs suppress noise but degrade resolution. To address this, we propose a hybrid coherence regulation (HCR) framework for high-quality imaging that computationally regulates the relative contributions of laser and LED amplitude maps following independent multidistance phase retrieval, achieving both resolution enhancement and noise suppression. On a hybrid-illumination lensless imaging system, this method achieves pixel-level resolution of 2.19 µm on a USAF 1951 target and reduces background noise variance. Furthermore, HCR produces H&E-like pseudocoloring from multiwavelength amplitude reconstructions of tissue sections across a reconstructed field of view (FOV) of 22.3 mm 2 , with color and structural fidelity surpassing traditional single-source illumination. This work provides an integrated framework for large field-of-view lensless imaging, offering high-throughput structure-preserving reconstruction and H&E-style color visualization.
- 3 November 2026
- Published by Optica Publishing Group in Optics Letters
- Vol. 51 (22) , 6179
High-quality lensless on-chip microscopy via mixed coherence control with H&E-like pseudocoloring
Achieving high-quality lensless microscopy remains challenging because of an inherent coherence trade-off: lasers enhance resolution but introduce artifacts, while LEDs suppress noise but degrade resolution. To address this, we propose a hybrid coherence regulation (HCR) framework for high-quality imaging that computationally regulates the relative contributions of laser and LED amplitude maps following independent multidistance phase retrieval, achieving both resolution enhancement and noise suppression. On a hybrid-illumination lensless imaging system, this method achieves pixel-level resolution of 2.19 µm on a USAF 1951 target and reduces background noise variance. Furthermore, HCR produces H&E-like pseudocoloring from multiwavelength amplitude reconstructions of tissue sections across a reconstructed field of view (FOV) of 22.3 mm 2 , with color and structural fidelity surpassing traditional single-source illumination. This work provides an integrated framework for large field-of-view lensless imaging, offering high-throughput structure-preserving reconstruction and H&E-style color visualization.
- Open Access
- 17 October 2026
- Published by Elsevier in SSRN Electronic Journal
QAD-Net: A Quality-Aware Dominant-Modality Network for Robust Multimodal Sentiment Analysis
Multimodal sentiment analysis integrates textual, acoustic, and visual cues to infer human sentiment, with practical applications requiring reliable predictions from asynchronous and incomplete inputs. However, unequal temporal resolutions and missing observations complicate cross-modal correspondence learning and weaken textual guidance, making it challenging to preserve sentiment-relevant information when modality availability varies. To address this challenge, this study proposes the Quality-Aware Dominant-Modality Network (QAD-Net), which combines CTC-inspired soft temporal aggregation, quality-aware mixture-of-experts proxy generation, adaptive text–proxy interpolation, and dominant-guided multimodal fusion; under random missing rates from 0.0 to 0.9, QAD-Net achieves average mean absolute errors of 1.032, 0.664, and 0.500 on CMU-MOSI, CMU-MOSEI, and CH-SIMS, respectively, while improving average non-zero binary accuracy on CMU-MOSI from 72.64% to 74.32% and average five-class accuracy on CH-SIMS from 34.51% to 38.71% over the strongest compared baselines on these metrics, demonstrating improved regression and classification performance under incomplete inputs. These findings support combining temporal aggregation with input-dependent compensation for robust sentiment prediction and indicate potential value for affective computing systems operating with asynchronous and partially missing multimodal observations.
- 17 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessQAD-Net: A Quality-Aware Dominant-Modality Network for Robust Multimodal Sentiment Analysis
Multimodal sentiment analysis integrates textual, acoustic, and visual cues to infer human sentiment, with practical applications requiring reliable predictions from asynchronous and incomplete inputs. However, unequal temporal resolutions and missing observations complicate cross-modal correspondence learning and weaken textual guidance, making it challenging to preserve sentiment-relevant information when modality availability varies. To address this challenge, this study proposes the Quality-Aware Dominant-Modality Network (QAD-Net), which combines CTC-inspired soft temporal aggregation, quality-aware mixture-of-experts proxy generation, adaptive text–proxy interpolation, and dominant-guided multimodal fusion; under random missing rates from 0.0 to 0.9, QAD-Net achieves average mean absolute errors of 1.032, 0.664, and 0.500 on CMU-MOSI, CMU-MOSEI, and CH-SIMS, respectively, while improving average non-zero binary accuracy on CMU-MOSI from 72.64% to 74.32% and average five-class accuracy on CH-SIMS from 34.51% to 38.71% over the strongest compared baselines on these metrics, demonstrating improved regression and classification performance under incomplete inputs. These findings support combining temporal aggregation with input-dependent compensation for robust sentiment prediction and indicate potential value for affective computing systems operating with asynchronous and partially missing multimodal observations.
- 6 October 2026
- Published by Optica Publishing Group in Optics Letters
- Vol. 51 (20) , 5713-5716
Single-shot non-invasive imaging through highly dynamic scattering media via autocorrelation-domain deconvolution
Imaging through dynamic scattering media remains challenging, because speckle patterns decorrelate during camera exposure. Here, we present a physics-based single-shot non-invasive imaging framework for dynamic scattering media. Enhancing speckle correlation imaging, the dynamically blurred autocorrelation is restored through autocorrelation-domain deconvolution using an analytical blur kernel. The restored autocorrelation is subsequently used for object recovery, with phase retrieval or a physics-informed network under severe degradation. Experimental results on dynamic ground glass and biological slide samples demonstrate robust object reconstruction under dynamic scattering conditions where conventional methods fail.
- 6 October 2026
- Published by Optica Publishing Group in Optics Letters
- Vol. 51 (20) , 5713-5716
Single-shot non-invasive imaging through highly dynamic scattering media via autocorrelation-domain deconvolution
Imaging through dynamic scattering media remains challenging, because speckle patterns decorrelate during camera exposure. Here, we present a physics-based single-shot non-invasive imaging framework for dynamic scattering media. Enhancing speckle correlation imaging, the dynamically blurred autocorrelation is restored through autocorrelation-domain deconvolution using an analytical blur kernel. The restored autocorrelation is subsequently used for object recovery, with phase retrieval or a physics-informed network under severe degradation. Experimental results on dynamic ground glass and biological slide samples demonstrate robust object reconstruction under dynamic scattering conditions where conventional methods fail.
- 21 September 2026
- Published by Optica Publishing Group in Optics Letters
- Vol. 51 (19) , 5373-5376
Symmetric-anchored transport-of-intensity phase retrieval with adaptive gradient refinement
Quantitative phase imaging via the transport-of-intensity equation (TIE) enables label-free phase measurement, but its accuracy degrades under large defocus, where the underlying nonlinearity becomes pronounced. We propose SA-RAdam, a symmetric-anchored, rectified-Adam-refined framework for high-fidelity TIE phase retrieval. A symmetric two-plane architecture yields a second-order accurate derivative prior at a virtual mid-plane, anchored to the reference plane via angular-spectrum propagation. An adjoint gradient of the forward model, combined with rectified adaptive momentum and line search, then recovers nonlinear high-frequency phase details typically lost in linear solvers. Simulations and experiments on binary and biological phase targets show consistent convergence, elimination of boundary overshoot, and RMSE reductions of over 60% relative to existing iterative solvers, substantially extending the practical defocus range of TIE-based phase retrieval.
- 21 September 2026
- Published by Optica Publishing Group in Optics Letters
- Vol. 51 (19) , 5373-5376
Symmetric-anchored transport-of-intensity phase retrieval with adaptive gradient refinement
Quantitative phase imaging via the transport-of-intensity equation (TIE) enables label-free phase measurement, but its accuracy degrades under large defocus, where the underlying nonlinearity becomes pronounced. We propose SA-RAdam, a symmetric-anchored, rectified-Adam-refined framework for high-fidelity TIE phase retrieval. A symmetric two-plane architecture yields a second-order accurate derivative prior at a virtual mid-plane, anchored to the reference plane via angular-spectrum propagation. An adjoint gradient of the forward model, combined with rectified adaptive momentum and line search, then recovers nonlinear high-frequency phase details typically lost in linear solvers. Simulations and experiments on binary and biological phase targets show consistent convergence, elimination of boundary overshoot, and RMSE reductions of over 60% relative to existing iterative solvers, substantially extending the practical defocus range of TIE-based phase retrieval.
- 20 July 2026
- Published by AIP Publishing in Applied Physics Letters
- Vol. 129 (3)
High dynamic range lensless on-chip microscopy via an overexposure inverse-projection operator
Lensless on-chip microscopy (LoM) enables high-throughput, label-free imaging with a compact setup and high numerical aperture. Its performance is limited by sensor saturation, which distorts high-intensity measurements and degrades reconstruction of the complex field. We report a high-dynamic-range inverse-projection reconstruction (HDR-IPR) framework that models saturation via an overexposure operator and jointly reconstructs the complex object field without multiple acquisitions at each predefined exposure duration. Using an approximate inverse-projection operator and alternating direction projection with complex total variation regularization, HDR-IPR effectively restores saturated signals while suppressing noise. Experiments demonstrate superior reconstruction quality, cleaner backgrounds, and enhanced reconstruction of fine structural details, providing a practical solution for HDR lensless imaging with standard sensors.
- 20 July 2026
- Published by AIP Publishing in Applied Physics Letters
- Vol. 129 (3)
High dynamic range lensless on-chip microscopy via an overexposure inverse-projection operator
Lensless on-chip microscopy (LoM) enables high-throughput, label-free imaging with a compact setup and high numerical aperture. Its performance is limited by sensor saturation, which distorts high-intensity measurements and degrades reconstruction of the complex field. We report a high-dynamic-range inverse-projection reconstruction (HDR-IPR) framework that models saturation via an overexposure operator and jointly reconstructs the complex object field without multiple acquisitions at each predefined exposure duration. Using an approximate inverse-projection operator and alternating direction projection with complex total variation regularization, HDR-IPR effectively restores saturated signals while suppressing noise. Experiments demonstrate superior reconstruction quality, cleaner backgrounds, and enhanced reconstruction of fine structural details, providing a practical solution for HDR lensless imaging with standard sensors.
- Open Access
- 18 July 2026
- Published by Elsevier in SSRN Electronic Journal
CARF: Reliability-Aware Adaptive Semantic Anchoring for Incomplete Multimodal Sentiment Analysis
Existing methods for incomplete multimodal sentiment analysis mainly focus on recovering missing features or aligning incomplete representations. However, many of them implicitly assume a fixed cross-modal reference structure, although the most reliable modality may change with the missing pattern and missing severity. This mismatch makes multimodal fusion unstable when the semantic anchor of cross-modal interaction shifts under incomplete conditions. To address this issue, we propose CARF, a reliability-aware adaptive semantic anchoring framework for incomplete multimodal sentiment analysis. CARF estimates modality reliability from representation statistics and observed missingness, constructs a soft semantic anchor in a shared space, and performs anchor-guided cross-modal interaction around the adaptive anchor. In addition, CARF calibrates observed, fused, and reconstructed evidence with missingness-conditioned reliability coefficients and uses a progressive missing-rate schedule during training. Experiments on CMU-MOSI, CMU-MOSEI, and CH-SIMS show that CARF improves classification and regression performance under multiple missing-modality settings. Ablation and anchor-selection studies further show that adaptive semantic anchoring and reliability-aware evidence calibration contribute to more stable multimodal reasoning when modalities are incomplete.
- 18 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCARF: Reliability-Aware Adaptive Semantic Anchoring for Incomplete Multimodal Sentiment Analysis
Existing methods for incomplete multimodal sentiment analysis mainly focus on recovering missing features or aligning incomplete representations. However, many of them implicitly assume a fixed cross-modal reference structure, although the most reliable modality may change with the missing pattern and missing severity. This mismatch makes multimodal fusion unstable when the semantic anchor of cross-modal interaction shifts under incomplete conditions. To address this issue, we propose CARF, a reliability-aware adaptive semantic anchoring framework for incomplete multimodal sentiment analysis. CARF estimates modality reliability from representation statistics and observed missingness, constructs a soft semantic anchor in a shared space, and performs anchor-guided cross-modal interaction around the adaptive anchor. In addition, CARF calibrates observed, fused, and reconstructed evidence with missingness-conditioned reliability coefficients and uses a progressive missing-rate schedule during training. Experiments on CMU-MOSI, CMU-MOSEI, and CH-SIMS show that CARF improves classification and regression performance under multiple missing-modality settings. Ablation and anchor-selection studies further show that adaptive semantic anchoring and reliability-aware evidence calibration contribute to more stable multimodal reasoning when modalities are incomplete.
- Open Access
- 9 July 2026
- Published by Elsevier in SSRN Electronic Journal
CDMR: Completeness-Aware Dominant Modality Refinement for Robust Multimodal Sentiment Analysis
Multimodal sentiment analysis under incomplete data aims to infer sentiment polarity when one or more modalities are missing, corrupted, or unreliable. This setting better reflects real-world applications than fully observed multimodal scenarios. Although existing methods have improved the robustness of multimodal sentiment analysis under incomplete conditions, they still have limitations in modeling missing uncertainty and exploiting the dominant role of different modalities. In most sentiment analysis tasks, the text modality usually provides more explicit sentiment cues than visual and audio modalities. However, incomplete multimodal inputs may weaken the reliability of textual representations. To address these issues, we propose a Completeness-Aware Dominant Modality Refinement(CDMR) framework for robust multimodal sentiment analysis under incomplete data. The framework first introduces a Dominant Modality Refinement Generator(DRG) to refine the text-dominant representation. Specifically, a completeness-aware weighting mechanism is designed to estimate the reliability of textual information and integrate original textual features with cross-modal reconstructed information. Then, we construct a hierarchical global-local semantic interaction structure to capture complementary multimodal information. The Global Multimodal Aggregation (GMA) module models high-level semantic consistency across modalities, while the Local Cross-modal Alignment(LCA) module captures fine-grained cross-modal relationships. An adaptive fusion mechanism is then employed to balance global and local representations under different missing scenarios. Finally, the multi-granular multimodal features are fused for sentiment polarity prediction. Various experiments on the MOSI and MOSEI datasets demonstrate the effectiveness of the proposed framework.
- 9 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCDMR: Completeness-Aware Dominant Modality Refinement for Robust Multimodal Sentiment Analysis
Multimodal sentiment analysis under incomplete data aims to infer sentiment polarity when one or more modalities are missing, corrupted, or unreliable. This setting better reflects real-world applications than fully observed multimodal scenarios. Although existing methods have improved the robustness of multimodal sentiment analysis under incomplete conditions, they still have limitations in modeling missing uncertainty and exploiting the dominant role of different modalities. In most sentiment analysis tasks, the text modality usually provides more explicit sentiment cues than visual and audio modalities. However, incomplete multimodal inputs may weaken the reliability of textual representations. To address these issues, we propose a Completeness-Aware Dominant Modality Refinement(CDMR) framework for robust multimodal sentiment analysis under incomplete data. The framework first introduces a Dominant Modality Refinement Generator(DRG) to refine the text-dominant representation. Specifically, a completeness-aware weighting mechanism is designed to estimate the reliability of textual information and integrate original textual features with cross-modal reconstructed information. Then, we construct a hierarchical global-local semantic interaction structure to capture complementary multimodal information. The Global Multimodal Aggregation (GMA) module models high-level semantic consistency across modalities, while the Local Cross-modal Alignment(LCA) module captures fine-grained cross-modal relationships. An adaptive fusion mechanism is then employed to balance global and local representations under different missing scenarios. Finally, the multi-granular multimodal features are fused for sentiment polarity prediction. Various experiments on the MOSI and MOSEI datasets demonstrate the effectiveness of the proposed framework.
- Open Access
- 1 July 2026
- Published by Elsevier in SSRN Electronic Journal
PCAMP: Personality-Conditioned Anchored Multimodal Perception for Robust Sentiment Analysis
Robust multimodal sentiment analysis (MSA) must handle acoustic and visual cues that are frequently noisy, exaggerated, or misleading. Most existing fusion methods integrate evidence from all modalities indiscriminately, without asking whether each signal should be trusted. We propose PCAMP, a personality-guided consistency-regularization framework that uses personality supervision as a trait-conditioned signal for regulating multimodal evidence before, during, and after fusion. The Personality-Guided Transformer (PGT) acts before fusion: it decomposes acoustic and visual features into anchor-aligned and anchor-orthogonal components, then uses a deviation-aware soft gate to attenuate unreliable acoustic-visual signals. A personality-conditioned fusion gate then adaptively weights text, acoustic, and visual summaries. Finally, the Personality-Modulated Alignment (PMA) module regularizes the fused representation toward a personality-conditioned stabilization target. On CMU-MOSI and CMU-MOSEI, PCAMP attains 87.19% and 87.63% binary accuracy under the include-zero protocol, and 88.81% and 88.97% under the exclude-zero protocol. Paired bootstrap comparisons against a matched personality-enhanced baseline show statistically supported gains. Diagnostic analyses further indicate that PCAMP retains its advantage under controlled audio-visual stress and that these improvements are not reducible to conventional personality-as-feature fusion. The results suggest that personality supervision, when used as a conditioning signal rather than a fused feature, offers a principled path toward more reliable multimodal integration.
- 1 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPCAMP: Personality-Conditioned Anchored Multimodal Perception for Robust Sentiment Analysis
Robust multimodal sentiment analysis (MSA) must handle acoustic and visual cues that are frequently noisy, exaggerated, or misleading. Most existing fusion methods integrate evidence from all modalities indiscriminately, without asking whether each signal should be trusted. We propose PCAMP, a personality-guided consistency-regularization framework that uses personality supervision as a trait-conditioned signal for regulating multimodal evidence before, during, and after fusion. The Personality-Guided Transformer (PGT) acts before fusion: it decomposes acoustic and visual features into anchor-aligned and anchor-orthogonal components, then uses a deviation-aware soft gate to attenuate unreliable acoustic-visual signals. A personality-conditioned fusion gate then adaptively weights text, acoustic, and visual summaries. Finally, the Personality-Modulated Alignment (PMA) module regularizes the fused representation toward a personality-conditioned stabilization target. On CMU-MOSI and CMU-MOSEI, PCAMP attains 87.19% and 87.63% binary accuracy under the include-zero protocol, and 88.81% and 88.97% under the exclude-zero protocol. Paired bootstrap comparisons against a matched personality-enhanced baseline show statistically supported gains. Diagnostic analyses further indicate that PCAMP retains its advantage under controlled audio-visual stress and that these improvements are not reducible to conventional personality-as-feature fusion. The results suggest that personality supervision, when used as a conditioning signal rather than a fused feature, offers a principled path toward more reliable multimodal integration.
- Open Access
- 30 June 2026
- Published by Elsevier in SSRN Electronic Journal
Robust Multimodal Sentiment Analysis under Missing Modalities via Hierarchical Interaction and Differentiated Reconstruction
Multimodal sentiment analysis (MSA) has become a crucial component of affective computing and multimodal video understanding, yet its reliability in real-world applications is severely compromised by inevitable data corruption and missing modalities (e.g., visual occlusion or frame loss). Current approaches exhibit limited cross-modal interaction, struggle to fully capture hierarchical dependencies, and often employ simplistic integrations of global and local features. Furthermore, conventional unified reconstruction strategies may cause irreversible semantic loss, significantly degrading model robustness under modality-deficient conditions. To address these critical challenges, we propose a novel framework: the Hierarchical Cross-Modal Transformer with Differentiated Modality Reconstruction (HCT-DMR). First, the HCT achieves deep fusion of multimodal features through a synergistic mechanism of hierarchical interaction and global semantic integration, effectively bridging modality gaps. Second, the DMR module applies mathematical constraints tailored to the distinct statistical distributions of text, audio, and video modalities (e.g., non-negative constraints for visual facial action units). This guarantees that the reconstructed features strictly align with their original semantic spaces, preventing representation degradation. Additionally, we introduce an Enhanced Temporal Encoding Architecture (ETEA) to strengthen non-verbal feature representation in complex temporal environments. Extensive experiments on three benchmark datasets (CMU-MOSI, CMU-MOSEI, and CH-SIMS) demonstrate that HCT-DMR establishes state-of-the-art robustness, providing a highly reliable sentiment prediction framework for complex, modality-deficient visual-linguistic scenarios.
- 30 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessRobust Multimodal Sentiment Analysis under Missing Modalities via Hierarchical Interaction and Differentiated Reconstruction
Multimodal sentiment analysis (MSA) has become a crucial component of affective computing and multimodal video understanding, yet its reliability in real-world applications is severely compromised by inevitable data corruption and missing modalities (e.g., visual occlusion or frame loss). Current approaches exhibit limited cross-modal interaction, struggle to fully capture hierarchical dependencies, and often employ simplistic integrations of global and local features. Furthermore, conventional unified reconstruction strategies may cause irreversible semantic loss, significantly degrading model robustness under modality-deficient conditions. To address these critical challenges, we propose a novel framework: the Hierarchical Cross-Modal Transformer with Differentiated Modality Reconstruction (HCT-DMR). First, the HCT achieves deep fusion of multimodal features through a synergistic mechanism of hierarchical interaction and global semantic integration, effectively bridging modality gaps. Second, the DMR module applies mathematical constraints tailored to the distinct statistical distributions of text, audio, and video modalities (e.g., non-negative constraints for visual facial action units). This guarantees that the reconstructed features strictly align with their original semantic spaces, preventing representation degradation. Additionally, we introduce an Enhanced Temporal Encoding Architecture (ETEA) to strengthen non-verbal feature representation in complex temporal environments. Extensive experiments on three benchmark datasets (CMU-MOSI, CMU-MOSEI, and CH-SIMS) demonstrate that HCT-DMR establishes state-of-the-art robustness, providing a highly reliable sentiment prediction framework for complex, modality-deficient visual-linguistic scenarios.