DYDongnan Yang

Dongnan Yang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Dongnan Yang has a total of 12 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    VASMA: Manifold-Guided Geometric Information Fusion for Bridging the Modality Gap

    by Hanwei Zeng, Lina Yang, Thomas Wu, Miao Zhou, Haoyan Yang, Dongnan Yang

    Zero- and few-shot recognition is bottlenecked by the modality gap---a systematic misalignment where textual anchors fail to match the visual manifold. We present VASMA, a Manifold-Guided Geometric Information Fusion framework bridging this topological void by enforcing strict geometric consistency across heterogeneous evidence sources.Transcending mere data expansion, VASMA utilizes LLM-expanded prompts as shared semantic anchors to orchestrate dual-domain evidence accumulation. It synergizes Pixel-Level coverage with Latent Manifold fidelity, ensuring comprehensive decision space exploration. Crucially, a tangent-space projection mechanism acts as a geometric filter, strictly aligning synthetic features with the intrinsic visual distribution. Heterogeneous streams are integrated into a Unified Cache, enabling auditable and calibrated fusion via a shared query kernel.VASMA establishes state-of-the-art results across 11 datasets, surpassing zero-shot CLIP by +24.6\% while achieving 73.9\% few-shot accuracy (+15.0 point gain). These findings substantiate that correcting the anchor-real geometric mismatch is prerequisite for robust multi-source fusion in data-scarce regimes.

  • Similarity-Guided Denoising Reconstruction for Unsupervised Image Captioning

    by Dongnan Yang, Lina Yang, Thomas Wu, Xichun Li, Yuan Yan Tang, Patrick Shen-Pei Wang

    Image captioning aims to generate natural and accurate textual descriptions of given images. Although significant progress has been made in image captioning models in recent years, most existing approaches heavily rely on high-quality image-text paired datasets that require expensive human annotation, thus limiting model scalability. Current unsupervised image captioning methods primarily focus on leveraging zero-shot learning capabilities of large pre-trained models (e.g. CLIP, GPT-2), yet still face persistent challenges including modality gaps, inefficient inference, and excessive noise incorporation, which constrain model accuracy and generalization capabilities. To address these limitations, we propose SGDR-Cap ( Similarity- Guided Denoising Reconstruction for Captioning), a novel unsupervised image captioning method that bridges the vision-language modality gap through a similarity-guided denoising reconstruction module. Our method leverages similarity information to guide the reconstruction of authentic text features during caption generation while simultaneously forcing the model to learn how to extract crucial image-relevant features and filter out unnecessary noise information. This enhances both coarse- and fine-grained cross-modal alignment. Furthermore, our approach jointly optimizes denoising reconstruction loss and language modeling loss, ensuring accuracy and fluency, and promoting greater diversity. Extensive evaluations on the MSCOCO and Flickr30K benchmarks demonstrate that our method achieves state-of-the-art results across all major metrics, with the most notable gain on the CIDEr score, improving from 101.1 to 104.4.

  • Emo-DiT: Emotional Speech Synthesis With a Diffusion Model Approach to Enhance Naturalness and Emotional Expressiveness

    by Yao Wang, Lina Yang, Bingzhen Wang, Jinhong Xu, Dongnan Yang, Miao Zhou, Yuan Yan Tang

    Current emotional text-to-speech tasks have achieved high-quality emotional speech by incorporating emotion modules into text-to-speech models. However, there has been limited in-depth research on embedding emotion modules within TTS models, and the expression of emotion in synthesized speech is constrained by both the TTS model and the emotion module, often preventing optimal results. This paper presents a novel TTS model, Grad-DiT, based on the DiT architecture of diffusion models, aimed at enhancing the naturalness and expressiveness of TTS. Unlike traditional U-net architectures, Grad-DiT leverages the Transformer architecture to better capture contextual information in text, resulting in more natural speech generation. Building on this model, we propose Emo-DiT, which incorporates an Emotion Feature Reconstruction (EFR) module to enable the synthesis of speech with specific emotional expressions. Experimental results show that Grad-DiT not only surpasses existing TTS models in speech quality but also significantly improves real-time performance and inference speed. Compared to traditional emotion generation methods, Emo-DiT offers more precise emotional expression, enabling the synthesis of speech with distinctive emotional characteristics and providing substantial support for future applications in emotional speech synthesis.