Huanyu Cheng

Dr. Huanyu Cheng

orcid

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
4 March 2025 in IEEE Access
Verified
Journal Article
Open access
HCHuanyu Cheng
Huanyu Cheng
GYGu Yingcheng
Gu Yingcheng
XMXi Mengting
Xi Mengting
ZQZhong Qiuyuan
Zhong Qiuyuan
WLWei Liu
Wei Liu
Detecting collapsed transmission line towers in remote sensing images is of great importance for post-disaster reconstruction of transmission lines. However, the scarcity of such fault samples makes it challenging for mainstream deep learning-based object detection algorithms to perform effectively, leading to unsatisfactory detection results. Recently, large language model (LLM) and multi-modal large models (MMLM) have demonstrated strong zero-shot and few-shot detection capabilities. However, due to their reliance on text-based understanding of image content, they struggle with fine-grained image details, making it difficult to accurately detect the targets and locations of collapsed towers in remote sensing images. This paper proposes a hybrid method combining large vision model and our improved small model for few-shot detection of collapsed transmission towers. We firstly improve YOLO V10 method by replace its large kernel structure with our new structure called IKL. And then we propose a collapsed tower detection method which leverages the strong zero-shot and few-shot learning capabilities of MMLM, while utilizing the fine-grained localization and detailed description capabilities of our improved small model. This synergy overcomes the limitations of each model type while retaining their advantages. Experimental results demonstrate that the proposed method effectively distinguishes between collapsed and intact towers and achieves superior few-shot learning performance compared to using MMLM or small models alone, significantly improving the accuracy of collapse detection.

17 December 2025 in Computers
Verified
Journal Article
Open access
PDF file
HCHuanyu Cheng
Huanyu Cheng
LSLinjiang Shang
Linjiang Shang
XCXikang Chen
Xikang Chen
TFTao Feng
Tao Feng
YZYin Zhang
Yin Zhang
Building high-quality multimodal instruction datasets is often time-consuming and costly. Recent studies have shown that a small amount of carefully selected high-quality data can be more effective for improving LVLM performance than large volumes of low-quality data. Based on these observations, we propose an error-guided multimodal sample selection framework with hallucination suppression for LVLM fine-tuning. First, semantic embeddings of queries are clustered to form balanced subsets that preserve task diversity. A visual contrastive decoding module is then used to reduce hallucinations and expose genuinely difficult examples. For closed-ended tasks, such as object detection, we estimate sample value using prediction accuracy; for open-ended question answering, we use the perplexity of generated responses as a difficulty signal. Within each cluster, high-error or high-perplexity samples are preferentially selected to construct a compact yet informative training set. Experiments on the InsPLAD detection benchmark and the PowerQA visual question answering dataset show that our method consistently outperforms random sampling under the same data budget, achieving higher F1, cosine similarity, BLEU (Bilingual Evaluation Understudy), and GPT-4o-based evaluation scores. This demonstrates that hallucination-aware, uncertainty-driven data selection can improve LVLM robustness and data efficiency.