CZChenghua Zhu

Chenghua Zhu

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Chenghua Zhu has a total of 14 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    COMET: Contrastive Motion-Enhanced Temporal Reasoning for Video Multimodal Large Language Models

    by Chenghua Zhu, Zhaolu Kang, Qifan Shi, Siyan Wu, Kehan Jiang, Lei Wei, Lianyu Hu, Guangyuan Dong, Mingbo Yang, Rui Lu, et al.

  • Open Access

    D2FNet: A Lightweight Dual-Driven Texture–Semantic Fusion Network for Fine-Grained Real-Time UAV Weed–Crop Detection

    by Chenghua Zhu, Siyan Wu, Xingqi Chao, Minxuan Lao, Lei Yang, Lihua Cai

    Weed–crop object detection in UAV field imagery faces several significant challenges, including a large proportion of small objects, dense occlusions, similar texture appearance, and strong background interference. These challenges often lead to missed detections, localization drift, and unstable training under edge-device budget constraints. To improve detection accuracy while maintaining a practical accuracy–efficiency trade-off in complex farmland scenes, we propose the Dual-Driven Texture–Semantic Fusion Network (D2FNet), consisting of a Texture–Semantic Backbone (TSB), an efficient operator MCF-A2C2f, a cross-scale adaptive fusion and feature redistribution module DSSA-Head, and a scale-aware reweighting block PSBL. TSB reduces discriminative ambiguity caused by similar weed–crop appearance and complex background textures; MCF-A2C2f controls the additional cost of the dual-driven design via lightweight operator substitution while largely preserving per-scale representations; DSSA-Head addresses multi-scale representation inconsistency induced by abundant small objects and large scale variation in field scenes; PSBL downweights low-quality positives by sample quality to stabilize box regression and training. Experimental results show that on the WeedCrop Image Dataset, D2FNet-n improves mAP50–95 from 36.6% to 44.1% (+7.5%) over the baseline YOLOv12-n; on the auxiliary Sesame Crop & Weed Dataset, mAP50–95 increases from 62.2% to 70.1% (+7.9%). These results indicate that D2FNet achieves stable accuracy gains under comparable parameter and computation budgets, rather than pursuing the smallest absolute model size, and shows promising cross-dataset robustness on the evaluated benchmarks.