Xiangyu Cheng

Mr. Xiangyu Cheng

orcid
Status: Accepted & In Press, Early Access available online i...

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

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

Representative Publications

View All Publications
7 April 2026 in Biomimetics
Verified
Journal Article
Open access
PDF file
Xiangyu ChengXiangyu Cheng
Xiangyu Cheng
MZmin zhou
min zhou
LZLiping Zhang
Liping Zhang
ZZZhang Zikai
Zhang Zikai
Metaheuristic optimization algorithms have attracted considerable research interest for solving complex optimization problems, yet many existing algorithms suffer from premature convergence and an inadequate balance between exploration and exploitation. The Coati Optimization Algorithm (COA) is a recently proposed nature-inspired metaheuristic that models the hunting and escape behaviors of coatis; however, it exhibits limited search diversity and tends to stagnate in local optima on high-dimensional, multimodal landscapes. This paper proposes an Improved Coati Optimization Algorithm (ICOA) that integrates four complementary enhancement strategies: (1) a Dynamic Adaptive Step-Size strategy that combines Lévy flights with Student’s t-distribution perturbations for heavy-tailed exploration; (2) a Population-Adaptive Dynamic Perturbation strategy that incorporates differential evolution operators with fitness-proportional scaling; (3) an Iterative-Cyclic Differential Perturbation strategy that employs sinusoidal scheduling and population-differential guidance; and (4) a Cosine-Adaptive Gaussian Perturbation strategy for refined exploitation with time-decaying intensity. ICOA is evaluated on 29 CEC2017, 10 CEC2020, and 12 CEC2022 benchmark functions across dimensions ranging from 10 to 100, compared against seven state-of-the-art algorithms in each benchmark suite. A statistical analysis using the Friedman test and the Wilcoxon rank-sum test confirms that ICOA achieves overall rank 1 on all three benchmark suites, with Friedman mean ranks of 1.207 (CEC2017, D=100), 1.000 (CEC2020, D=10), and 2.208 (CEC2022, D=10); the CEC2020 result should be interpreted in the context of its low dimensionality. A scalability analysis across four dimensionalities (10D, 30D, 50D, 100D) demonstrates consistent first-place rankings with mean ranks between 1.000 and 1.207. An ablation study and a sensitivity analysis of the strategy activation probability validate the contribution of each individual strategy and the optimality of the 50% activation setting. Furthermore, ICOA achieves the best results on all six constrained engineering design problems tested, with all improvements confirmed as statistically significant (p<0.05).

24 August 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
XCXiangyu Cheng
Xiangyu Cheng
Xiangyu ChengXiangyu Cheng
Xiangyu Cheng
WZWeipeng Zeng
Weipeng Zeng
MZMeng Zha
Meng Zha
feng xiangfeng xiang
feng xiang
Real-time object detection from unmanned aerial vehicle (UAV) imagery is an important engineering requirement in power-line inspection, railway monitoring, traffic supervision, and maritime surveillance. It remains difficult because industrial aerial scenes contain dense small targets, large scale variation, motion-induced degradation, and cluttered backgrounds. This paper proposes a Cross-scale Multi-level Information Fusion Enhancement detector based on the You Only Look Once architecture (CMIFES-YOLO) for lightweight UAV-based industrial inspection. The artificial intelligence contribution is a cross-layer adaptive feature-fusion design that combines lightweight squeeze-and-excitation channel recalibration, residual-initialized spatial attention, and a Scale-Aware Cross-layer Reweighting mechanism to learn the relative importance of scale-diverse feature streams at each detection level. A dual-cascaded fusion structure is further deployed at the finest detection level to strengthen small-object representations, while single fusion modules at deeper levels improve medium- and large-scale perception. The engineering application is real-time aerial inspection on resource-constrained UAV platforms. Under a unified retraining protocol on the VisDrone benchmark, CMIFES-YOLO achieves 41.72\% mAP@0.5 and 25.35\% mAP@0.5:0.95, improving YOLOv11n by 6.74 and 5.55 percentage points, respectively. The detector contains 6.17 million parameters, requires 15.5 GFLOPs, and runs at 82 FPS on an RTX 4060 GPU, satisfying real-time deployment requirements. Cross-dataset evaluation on DUT-Anti-UAV further indicates positive transfer without fine-tuning. These results show that adaptive cross-scale feature fusion can improve UAV inspection accuracy while retaining practical inference efficiency.

16 October 2026 in Array
Verified
Journal Article
Open access
Xiangyu ChengXiangyu Cheng
Xiangyu Cheng
MZMin Zhou
Min Zhou
WZWeipeng Zeng
Weipeng Zeng
MZMeng Zha
Meng Zha
FXFeng Xiang
Feng Xiang