Intelligent perception systems are essential for precision agriculture, yet their deployment in agricultural field environments is significantly challenged by dense target occlusion and strict resource constraints on edge devices. To address this issue, this paper reviews lightweight object detection technologies from a problem-driven perspective, focusing on the interaction between occlusion-induced feature degradation and limited model capacity under real-world conditions. Unlike existing surveys that mainly summarize model evolution or application scenarios, this work presents a systematic review based on a unified analytical framework to examine how lightweight models compensate for feature loss caused by complex physical factors. Specifically, we analyze the mechanisms underlying feature degradation arising from morphological similarity, extreme scale variation, and dynamic environmental disturbances such as illumination changes and non-rigid deformation. Based on this analysis, recent advances in lightweight detection architectures are comparatively reviewed, including the YOLO series, real-time Transformers, and State Space Models, with an emphasis on their design trade-offs between computational efficiency and representation capability. In addition, key optimization strategies are discussed, such as multi-scale attention mechanisms and dynamic routing for adaptive computation allocation, as well as distribution-aware loss functions for improving localization robustness in densely occluded scenarios. The role of large vision models is also explored, highlighting their lightweight adaptation through knowledge distillation and parameter-efficient fine-tuning. Overall, by synthesizing empirical findings and comparative evidence from the recent literature, this review provides a structured understanding of collaborative optimization pathways and offers evidence-based strategic insights into achieving an effective balance between detection accuracy and computational efficiency for agricultural edge deployment.
Yan et al. (Wed,) studied this question.