ZHZeliang Huang

Zeliang Huang

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zeliang Huang has a total of 10 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • FESA-ViT: Frequency-Enhanced Staged Attention Vision Transformer With Decoupled Training for UAV Insulator Inspection

    by Zeliang Huang, Zhenyu Wang, Chunlei Li

    UAV power line inspection typically faces several formidable challenges, including cluttered background textures, drastic scale variations, and constrained edge computing resources. Although single-head attention Vision Transformers reduce complexity, their simplified interaction mechanisms struggle to handle the scale disparities between insulators and tiny defects. Furthermore, high-level semantics are easily degraded by complex background noise. To address these issues, we propose the frequency enhanced staged attention vision transformer (FESA-ViT). First, we design the single-head frequency enhanced staged attention block (SFESA Block), which integrates a partial channel token mixing strategy to reduce computational redundancy and embeds multiscale dilated convolutions to enhance cross-scale perception, while adopting a staged attention mechanism to balance the fine-grained extraction of tiny insulator defects and the modeling of long-range dependencies. Second, a spectrum token mixer (STM) is introduced in the deep layers to decouple background noise from semantic information via frequency-domain modulation, improving representation purity. Finally, we propose a training-inference decoupling strategy. The STM is active only during training and completely removed during inference, delivering significant performance gains over the STM-free baseline with zero extra computational cost. Experimental results show FESA-ViT achieves 62.3% and 46.0% mAP on the PTL-AI Furnas Dataset (D1) and Insulator Defect Dataset (D2), respectively. Even with STM unloaded, the model retains improved performance over the STM-free baseline, reaching 74 FPS inference speed on D1 with a slight mAP drop to 61.4%, thus outperforming mainstream networks in complex power inspection tasks.


  • Compact spatial arrangement of multiple sparse filters for fabric defect detection

    by Zeliang Huang, Chunlei Li, Miao Yu, Shuwei Shi, Xinnan Shao

    Fabric defect detection is one of the hot topics in the textile industry, primarily due to occasional machine failures and repetitive errors on production lines leading to misalignment in warp and weft. Consequently, collecting defect samples is challenging, as these defects appear as subtle, linear shapes in horizontal or vertical orientations, resulting in sparse defect features with low contrast. This poses significant challenges to the model’s feature learning capabilities and detection performance. we developed a research platform for developing an automatic fabric detection system, focusing on collecting fabric defect images and developing/deploying detection algorithms. Utilizing this platform, a large number of fabric images were collected and annotated on the denim production line in textile enterprises, with the dataset being made publicly available for research within the field. Mainstream methods such as local attention or multiscale fusion have limited effectiveness in detecting defects in denim, which exhibit sparse features, varied morphologies, low contrast, and a lack of global features. This paper designs multiple sparse filters for feature extraction. Square filters are utilized to extract minute defect features in fabric images, while asymmetric filters are employed to capture linear defect characteristics, enhancing the model’s representation capabilities. Furthermore, a compact arrangement involving multiple sparse filters is proposed to enhance feature diversity in model learning and reduce the risk of overfitting. Experiments conducted on the dataset provided in this paper and the TIANCHI dataset demonstrate that the method proposed in this paper outperforms existing mainstream algorithms. The denim fabric dataset can be obtained on GitHub: https://github.com/h-z-liang/DEMIN_DATASET.git .