Review of Lightweight Research Based on YOLOv4 Model

DOI: https://doi.org/10.62517/jbdc.202601217

Author(s)

Yuchen Lin

Affiliation(s)

Computer Science and Technology, Jinshan College, Fujian Agriculture and Forestry University, Quanzhou, Fujian, China

Abstract

In recent years, the demand for the application of object detection models in resource-constrained scenarios has been growing rapidly. Although detection algorithms represented by YOLOv4 demonstrate remarkable advantages in detection accuracy, their high computational complexity restricts deployment on mobile and edge devices. This paper systematically summarizes the research progress of YOLOv4 lightweight technologies, covering traditional compression methods such as model pruning, parameter quantization and knowledge distillation, as well as lightweight network improvement schemes based on depthwise separable convolution, attention mechanisms and neural architecture search. Comparative experimental data verify that replacing the backbone network with GhostNet can reduce the model size by 67% and improve the inference speed by 2.3 times, while maintaining 91% of the accuracy of the original model. The study further explores engineering application cases of model lightweight technologies in real-time detection scenarios such as UAV inspection and vehicle-mounted systems. Finally, aiming at the feature loss problem caused by model compression, future research directions combining dynamic convolution and hardware co-optimization are proposed.

Keywords

YOLOv4 Model; Object Detection; Model Lightweighting; Model Pruning; Knowledge Distillation

References

[1] Li J Q. (2023). Research on post-disaster personnel detection algorithm based on deep learning[D]. Qinhuangdao: Yanshan University. https://doi.org/10.27440/d.cnki.gysdu.2023.002145. [2] Zhu H Y. (2024). Research on power target detection algorithm for inspection images of transmission lines[D]. Qingdao: Qingdao University of Science and Technology. https://doi.org/10.27264/d.cnki.gqdhc.2024.000893. [3] Chen Z L, Chen Y R, Li Q Z, Zhong R, Zhu Y K. Research on indirect monocular pose estimation method for spacecraft based on deep learning[J]. Flight Control and Detection, in press. [4] Zhang L H. (2022). Research on target detection application based on RK3399Pro and YOLO[D]. Beijing: University of Chinese Academy of Sciences (Institute of Optoelectronics, Chinese Academy of Sciences). https://doi.org/10.27543/d.cnki.gkgdk.2022.000046. [5] Zhang X H. (2025). Method of infrared-visible image fusion and target detection in complex sea surface scenes[D]. Guangzhou: Guangdong University of Technology. https://doi.org/10.27029/d.cnki.ggdgu.2025.003872. [6] Jiang F, Xu Z T, Cui D Y, Yao H W, Zhang H Y, Han Y. (2025). Application of large artificial intelligence models in the prevention and control of forest and grass pests[J]. World Forestry Research, 38(06):43-48. https://doi.org/10.13348/j.cnki.sjlyyj.2025.0101.y. [7] Feng K Y. (2022). Sparse optimization and search of deep neural network structures[D]. Xi'an: Xidian University. https://doi.org/10.27389/d.cnki.gxadu.2022.003449. [8] Wang X W, Li X, Zhu Z H, Qin T, Zhu H, Luo A Q. Multi-class weed detection model based on improved YOLO11n[J]. Journal of Tianjin University of Technology, in press. [9] Yang Y C. (2022). Research on model compression method based on quantization[D]. Guangzhou: South China University of Technology. https://doi.org/10.27151/d.cnki.ghnlu.2022.002479. [10] Wang Q, Wang Z Q, Liu H, Liang M, Zhang Y C, Lin Y. (2025). Evolution law, application efficiency and future prospect of YOLO series models in wheat ear detection[J]. Hubei Agricultural Sciences, 64(12):218-227. https://doi.org/10.14088/j.cnki.issn0439-8114.2025.12.037.