A comprehensive review of lightweight deep learning models for edge computing with future directions

The rapid development of the Internet of Things (IoT) has driven widespread demand for edge intelligence. However, resource-constrained edge devices struggle to support complex deep learning models, making lightweight deep learning models a key research direction. Despite significant progress in this field, a systematic review of its knowledge structure, evolutionary path, and future directions remains lacking. To this end, this paper, based on the PRISMA guidelines, screened 543 articles from the Web of Science (WoS) Core Collection and employed a combined bibliometric and systematic analysis approach to systematically analyze the trends, themes, and directions in this field. Research indicates that this field has experienced exponential growth in recent years. Model compression has been widely adopted as a foundational approach; neural architecture search (NAS) has become a mainstream paradigm for automated design of efficient models; knowledge distillation (KD) and lightweight Transformer architectures have also emerged as important frontiers. Research is shifting from “ex post” model compression to efficiency-focused native architecture design and algorithm-hardware co-optimization. However, this field still faces key challenges, including the accuracy-efficiency-robustness trade-off, high training resource overhead, and a lack of evaluation benchmarks. This study systematically reveals the field’s development trajectory, providing theoretical reference and guidance for future research.