A systematic literature review of lightweight YOLO models for object detection

Selection of literature databases

We selected two leading and widely recognized academic databases, Scopus and the Web of Science (WoS) Core Collection, as our primary sources for literature retrieval. The rationale is threefold:

• Extensive coverage: Scopus and WoS collectively index a broad set of peer-reviewed journals and proceedings from major publishers (including MDPI, Springer Nature, IEEE, Elsevier, Frontiers, Tech Science Press, SAGE, IET, among others), which supports a representative literature sample.

• Quality control: Both databases prioritize peer-reviewed publications, helping ensure academic rigor of included studies.

• Mitigation of publisher bias: Using multiple comprehensive databases reduces reliance on any single publisher platform and helps to minimize selection bias.

Search keywords and strategy

The process of developing our search query, from the initial research statement to the final keywords and synonyms, is illustrated in Fig. 3. Based on our research questions and initial scoping, we developed a search string combining object detection terms, model names, and lightweighting terms to maximize recall while maintaining relevance. The keyword categories included:

• Object detection related: object, target, image, subject, item

• Detection action related: detection, identification, recognition, observation, spotting, sensing

• Model name: YOLO

• Lightweight related: lightweight, optimize*, minimiz*, compact, resource-constrained, tiny, lite, mobile, efficient

The final search string applied (adapted to each database syntax) was:

• (“object detection” OR “target detection” OR “object recognition” OR “target recognition”) AND “YOLO” AND (“lightweight” OR “optimize*” OR “minimiz*” OR “compact” OR “resource-constrained” OR “tiny” OR “lite” OR “mobile” OR “efficient”)

• This search strategy underwent iterative refinement through pilot queries to balance comprehensiveness and precision.

Timeframe and document types

• Timeframe: We limited our search to publications between 2016-01-01 and 2025-01-01 to cover the evolution of the YOLO family and related lightweighting methods.

• Document types: To ensure comparability and peer review quality, we only included peer-reviewed journal articles and peer-reviewed conference articles published in English. Workshop articles, technical reports, and other non-peer-reviewed materials were excluded (see Inclusion/Exclusion Criteria below for details).

Inclusion and exclusion criteria

Inclusion criteria:

• 1.

  peer-reviewed journal articles and peer-reviewed conference articles published in English.

• 2.

  Publication date between 2016-01-01 and 2025-01-01.

• 3.

  Explicit focus on YOLO or YOLO-based object detection with attention to one or more lightweighting strategies (e.g., pruning, quantization, knowledge distillation, lightweight backbones, Neural Architecture Search).

• 4.

  Empirical evaluation reported (datasets and quantization performance metrics).

Exclusion criteria:

• 1.

  Non-peer-reviewed materials (e.g., preprints, workshop abstracts not subject to peer review, technical reports, and student theses).

• 2.

  Non-English publications.

• 3.

  Articles that do not present experimental results or do not have a clear focus on YOLO or lightweighting techniques.

• 4.

  Duplicate publications: when the same work appears in multiple venues (e.g., a conference article later extended to a journal article), the most complete peer-reviewed version was retained and the duplicate(s) were excluded.

Screening and selection procedure

The literature screening followed PRISMA 2020 guidelines (Page et al., 2021) to ensure transparency and reproducibility. Two reviewers independently performed title and abstract screening and then full-text eligibility assessment. Disagreements at any stage were resolved by discussion; when consensus could not be reached, a third reviewer made the final decision. After duplicate removal, the initial search returned 485 unique records. Title screening excluded 138 records; abstract screening excluded a further 140 records; full-text assessment led to exclusion of 104 additional records (mainly because the studies did not specifically focus on lightweight YOLO or lacked sufficient empirical data). The multi-stage screening resulted in 103 peer-reviewed journal and conference articles that met all inclusion criteria and were included for detailed analysis (PRISMA flow shown in Fig. 4).

Data extraction, labelling and statistics

Data extraction and topic annotation (technical categories, application domains, hardware platforms, etc.) were manually performed by two independent reviewers. For each included study, we recorded metadata (authors, year, journal, DOI, publisher), the YOLO variant used, lightweighting methods, datasets, reported performance metrics (e.g., mean Average Precision (mAP)), model size/FLOPs/FPS (if available), and the hardware platform used for deployment experiments. The percentage statistics reported in this review (e.g., “35% of studies used pruning”) were calculated as follows: (number of studies explicitly applying the technique among the 103 included studies) $÷$ 103. Label inconsistencies were resolved through discussion and consensus.

Quality assessment

All 103 peer-reviewed articles underwent a pre-defined quality assessment (QA) using a 12-item checklist (Table 2). Each criterion was scored as P (Present = 1), S (Sufficient/partially met = 0.5) or A (Absent = 0). QA scores were mapped to the review’s research questions to support transparent interpretation of the evidence base.

Comparative analysis of lightweight YOLO models

To provide a meaningful comparison with existing methods, we present a performance benchmark of SS-YOLOv8n against various prominent YOLO models, including other lightweight variants. The data, adapted from Fan et al. (2024), highlights the trade-offs between model size, computational complexity, and detection accuracy.

As shown in Table 9, SS-YOLOv8n demonstrates a remarkable balance between efficiency and accuracy. It achieves the smallest model size (2.3 MB), lowest Giga Floating-point Operations Per Second (GFLOPs) (3.1 G), and fewest parameters (0.88 M) among all compared models, making it highly suitable for resource-constrained environments. Despite its compact nature, SS-YOLOv8n maintains a competitive mAP50 of 0.799 and mAP50–95 of 0.510, outperforming several larger models like YOLOv5n/s, YOLOv8n, YOLOv8-mobilentv4, and YOLOv8-ghost in mAP50. Furthermore, its inference speed (128.9 FPS) is the highest, significantly surpassing the original YOLOv8n (112.5 FPS) and other lightweight variants. This superior performance highlights the success of SS-YOLOv8n’s integrated lightweight design strategies, which optimize the model from backbone to detection head and loss function.

The effectiveness of these individual lightweight techniques is further validated by ablation studies and comparisons of different pruning methods, as presented in Fan et al. (2024). Given pruning’s significant adoption rate (as highlighted in Fig. 6) and its direct impact on model size and computational efficiency, a deeper dive into its specific application within SS-YOLOv8n is particularly relevant. For instance, their ablation experiments (e.g., Table 4 in Fan et al. (2024)) show that each proposed module (Weighted Shuffle Convolutional Block 2f (WS-C2f), Spatial Context and Depth Head (SCDH), Probabilistic Intersection over Union 2 (PIoU2)) contributes incrementally to the overall performance improvement. Moreover, their comparison of pruning methods (e.g., Table 6 in Fan et al. (2024)) indicates that the chosen slimming pruning approach effectively reduces model size while maintaining or even enhancing accuracy, demonstrating its superiority over other methods like Lamp and Group-norm at similar pruning rates. These findings underscore the importance of a holistic approach to lightweight model design, combining architectural innovations with effective compression techniques. Ultimately, the success of SS-YOLOv8n serves as a compelling answer to RQ1, demonstrating that a strategic combination of architectural redesign, efficient module integration, and targeted optimization (such as pruning) can yield highly efficient YOLO models capable of superior performance in resource-constrained environments, effectively balancing accuracy with computational demands.

Main applications of lightweight YOLO models (RQ2A)

There are eight main application directions in the screened literature, as shown in Table 10 summarizing the main application areas of lightweight YOLO models in the literature.

Cross-domain generalization: Lightweight YOLO models significantly improve the adaptability on resource-constrained devices through attention mechanisms, model compression (e.g., pruning, quantization), and multi-scale feature fusion.

Real-time and accuracy balance: Most of the literature reduces computation while maintaining accuracy by replacing backbone networks (e.g., MobileNet, GhostNet) and optimizing loss functions (e.g., WIoU, Focal Loss).

Edge-Cloud collaboration: Cloud-edge collaboration frameworks (e.g., ECC+) for industrial inspection and UAV applications significantly reduce end-to-end latency (88.9%) to support large-scale real-time tasks.

Small Target Detection Optimization: For agriculture and UAV scenarios, dedicated modules (e.g., HEPAN, DySample) are introduced to improve small target detection capability, with up to 10% mAP improvement.

Lightweight YOLO models show a wide range of application potential in the fields of intelligent transportation, industrial detection, agricultural monitoring, public safety, UAV applications, edge computing, medical detection and infrastructure security. These models achieve real-time and resource efficiency while improving detection accuracy through techniques such as attention mechanism, model compression and multi-scale feature fusion. Cross-domain studies show that lightweight YOLO models have significant advantages in small target detection and complex scene adaptation.

Impact of data sets on model performance (RQ2B)

A total of 61 articles in the screened literature used multiple datasets, with a special focus on dataset diversity and task applicability. The characteristics of different datasets (e.g., data volume, category diversity, contextual complexity, etc.) have a significant impact on the performance of lightweight YOLO models. Ten representative YOLO literatures on data diversity will be shown in Table 11.

A critical discovery from this analysis is that the diversity of datasets directly dictates the specific lightweighting strategies employed and profoundly impacts the model’s generalization ability and applicability. For example, the complex field environment of the IP102 pest dataset (Yu et al., 2024b) demands higher robustness from the model, while low-light environment datasets (Vinoth & Sasikumar, 2024) rigorously test the model’s detection ability under extreme conditions. These examples reveal a crucial insight: the inherent challenges posed by specific dataset characteristics (e.g., fine-grained details in pest detection, noise in low-light, occlusions in fire) directly dictate the *type* and *degree* of lightweighting and architectural modifications required. By introducing a specific attention mechanism or optimization module, the model can better adapt to the unique characteristics of the dataset and improve the accuracy and efficiency of task completion. This further underscores the nuanced trade-offs discussed in RQ1; for instance, aggressive pruning might be detrimental for small target detection datasets, necessitating more sophisticated architectural designs or knowledge distillation.

To achieve real-time detection on edge devices, many studies have strategically reduced model parameters and computation. For example, Yu et al. (2024b) achieved a 70.2% reduction in parameters through structured pruning with only a 0.8% mAP decrease, demonstrating the effectiveness of intelligent lightweight design. Concurrently, the significant improvement in FPS (e.g., 15.99% in Lin, Yun & Zheng, 2024) further validates the critical role of lightweight optimization for real-time performance.

The contextual complexity inherent in certain datasets, such as smoke and background interference in forest fire datasets (Lei et al., 2024a; Lin, Yun & Zheng, 2024), poses significant challenges to model robustness. Here, the insight is that targeted architectural enhancements are paramount. By introducing mechanisms like coordinate attention (Liu et al., 2024a), models can more accurately capture target features and improve detection accuracy even in highly cluttered or obscured scenes.

For resource-constrained edge devices, the optimization of parameter count and computational efficiency is paramount. For instance, Boyle et al. (2024) achieved an impressive latency of 0.6 ms/frame and energy consumption of 31 $\mu$J/frame on a RISC-V microcontroller. This highlights that lightweighting, when coupled with hardware-aware design, enables efficient models for highly constrained Internet of Things (IoT) scenarios.

Overall Insights and Unresolved Challenges for RQ2

Existing studies demonstrate significant progress in optimizing lightweight YOLO models for target detection. A key discovery is that the interplay between application demands and dataset characteristics is a primary determinant in the selection and refinement of lightweight YOLO techniques. While efficiency is paramount, maintaining accuracy across diverse and challenging real-world data remains a central challenge. The combination of structured pruning and novel convolutional technologies has enabled significant improvements in inference efficiency while preserving detection performance, successfully applied in edge computing scenarios like drones and embedded devices.

However, current research still faces critical limitations:

• Insufficient cross-scene generalization: Model training is often confined to single domains, leading to insufficient generalization ability across varied scenes. This reveals a critical research gap: the need for lightweight models that are not only efficient but also inherently adaptive and resilient to real-world variability. Adaptive migration methods are crucial for future exploration.

• Robustness to dynamic interference: Studies on robustness under dynamic environmental interference are relatively weak. Future work should focus on enhancing real-time response capabilities by integrating dynamic network architectures that can adapt to changing conditions.

• Energy-efficiency and hardware co-design: The energy-efficiency constraints of edge devices necessitate further optimization of model-hardware co-design. This involves developing lightweight architectures that are intrinsically compatible with low-power hardware, moving beyond mere software-level compression.

• Standardized evaluation metrics: The variability of existing evaluation metrics across studies hinders comparability. Establishing unified standards is essential to accurately assess and compare the performance of lightweight models across different research efforts.

At the practical level, a scenario-oriented optimization strategy is recommended. This involves focusing on accuracy enhancement for specific tasks like agricultural automation (tiny target recognition) and prioritizing real-time performance for industrial scenarios. Customizing model structures based on hardware characteristics is also vital. Open community collaboration, including sharing lightweight models and industry-specific datasets, is essential to accelerate cross-domain application innovation. Ultimately, the continuous development of lightweight YOLO technology requires integrating algorithmic innovation, hardware adaptation, and domain-specific knowledge, with core breakthroughs focusing on generalization capability improvement and energy efficiency optimization to provide reliable support for practical deployment.

What hardware platforms are suitable for deploying lightweight YOLO models? (RQ3)

For new researchers entering the field of lightweight YOLO model deployment, understanding the interplay between model design and hardware capabilities is crucial. The selection of a suitable hardware platform is not merely about computational power; it’s a strategic decision based on a delicate balance of computational performance, power requirements, cost, and the specific demands of the application scenario. This section aims to provide valuable insights into the diverse hardware landscape for lightweight YOLO models.

Based on the filtered literature, Table 12 summarizes 11 key articles focusing on hardware deployment, offering a comprehensive overview of suitable platforms and their characteristics.

This classification highlights the diverse range of hardware platforms, each tailored to specific operational contexts. For new researchers, a key takeaway is that the choice of hardware dictates the feasible lightweighting strategies and the ultimate performance envelope.

• GPU-accelerated edge devices (e.g., NVIDIA Jetson Series): These platforms offer significant parallel processing capabilities, making them suitable for applications requiring higher throughput and more complex models than purely CPU-based solutions. The insight here is that while they are “edge,” they still benefit immensely from software optimizations like TensorRT acceleration and model quantization, which maximize the utilization of their powerful GPUs. For researchers, this means focusing on GPU-friendly lightweight architectures and leveraging vendor-specific optimization tools.

• Low-cost general-purpose boards (e.g., Raspberry Pi): These are excellent starting points for budget-constrained projects or proof-of-concept deployments. The primary insight is that their CPU-centric architecture necessitates extreme model lightweighting (e.g., using MobileNet or Ghost modules as backbones) to achieve acceptable real-time performance. Researchers should be prepared for lower FPS compared to GPU-accelerated options and prioritize models with minimal computational demands.

• Ultra-low power microcontrollers (e.g., STM32, RISC-V MCUs, ARM Cortex-M Series): These platforms are designed for highly constrained environments where power consumption is paramount (e.g., battery-powered IoT devices, embedded systems). The critical insight for new researchers is that deployment here requires aggressive quantization (e.g., to INT8 or even INT4), highly optimized inference engines (like TFLite Micro, CMSIS-NN), and meticulous memory management. Performance metrics shift from FPS to “ms/inference” and “ $\mu$J/inference,” emphasizing energy efficiency per operation. This domain often involves a deeper understanding of hardware-software co-design.

• Field-programmable gate arrays (FPGAs): FPGAs offer unparalleled flexibility and energy efficiency for specific, custom AI workloads. The key insight is that FPGAs allow for hardware-level parallelism and custom pipeline designs, enabling highly optimized inference for fixed models. However, this comes at the cost of higher development complexity and specialized hardware description language (HDL) knowledge. For researchers, FPGAs are a powerful option when off-the-shelf solutions cannot meet stringent power or latency requirements.

• Cloud-edge collaboration systems: This represents a deployment strategy rather than a single hardware platform. The valuable insight is that not all AI processing needs to occur solely on the edge device. By intelligently distributing tasks between the edge (for immediate, real-time inference) and the cloud (for complex analysis, model updates, or large-scale data processing), this approach can overcome individual hardware limitations. Researchers should consider this hybrid model for applications requiring both low latency and extensive computational resources.

Key challenges and unresolved issues for new researchers (RQ3)

While lightweight YOLO models offer immense potential, their deployment on diverse hardware platforms presents significant challenges that new researchers must navigate:

• 1.

  Hardware-software ecosystem fragmentation: Different hardware vendors offer distinct SDKs and optimized libraries. This fragmentation means models optimized for one platform may not run efficiently on another without substantial porting and re-optimization, leading to a steep learning curve and potential vendor lock-in.

• 2.

  Resource limitations vs performance trade-offs: Strict hardware constraints (computational power, memory, power consumption) necessitate difficult trade-offs between model accuracy, inference speed, and energy efficiency. Achieving optimal utilization on constrained hardware requires meticulous model design and inference engine tuning.

• 3.

  Performance variability and benchmarking challenges: Real-world performance can vary significantly even on theoretically similar hardware due to differences in software stacks and optimizations. This makes accurate and reproducible benchmarking a complex task for new researchers.

• 4.

  Lack of unified toolchains for co-design: The absence of truly unified, end-to-end toolchains for model training, lightweighting, and hardware-aware deployment often forces researchers to integrate disparate tools and workflows.

To successfully deploy lightweight YOLO models, new researchers should adopt a holistic and scenario-driven approach:

• Definition constraints first: Clearly definition application constraints (latency, power, cost, accuracy) before hardware selection. These guide subsequent model optimization.

• Understand the hardware-software stack: Recognize that performance is a product of both model architecture and the underlying hardware-software stack (compilers, inference engines, libraries).

• Embrace hardware-aware lightweighting: Tailor lightweighting techniques (e.g., pruning, quantization) to the specific hardware, as effectiveness varies across platforms.

• Prioritize real-world benchmarking: Conduct practical benchmarks on target hardware to measure actual performance (FPS, latency, power) under realistic conditions, rather than relying solely on theoretical metrics.

• Consider hybrid approaches: For complex applications, explore cloud-edge collaboration or heterogeneous computing to leverage the strengths of different processing units.

• Leverage community resources: Utilize open-source pre-trained models and deployment examples to accelerate development and learn from existing implementations.

In essence, deploying lightweight YOLO models is a multidisciplinary endeavor demanding a deep understanding of both AI principles and hardware characteristics. New researchers must navigate a complex ecosystem, make informed trade-offs, and continuously optimize to meet real-world demands. The field’s future hinges on bridging the software-hardware gap for more seamless and efficient AI deployment.

What performance metrics are used to evaluate lightweight YOLO models? (RQ4)

For researchers, especially those new to the field, understanding the appropriate performance metrics is paramount for effectively evaluating and comparing lightweight YOLO models. The evaluation of these models is inherently multi-faceted, requiring a comprehensive set of metrics that capture both their accuracy and their efficiency. This section delves into the standard evaluation metrics, their significance, and provides insights into how researchers can best utilize them.

The performance evaluation of lightweight YOLO models typically encompasses several key metrics, including accuracy-related measures (e.g., mAP, precision, recall) and efficiency-related measures (e.g., inference speed, model size, computational complexity). Together, these metrics form a comprehensive evaluation system to measure the overall performance of the model in different scenarios. Table 13 showcases a few randomly selected research articles from the screened literature, illustrating the consistency in the core performance metrics used for experimental comparison.

As observed, the core performance metrics consistently employed within the lightweighting literature include mAP50–95, precision, recall, and GFLOPs. However, a truly comprehensive evaluation also considers other crucial efficiency metrics like model parameters and actual inference speed (FPS/Latency). These key performance metrics are outlined in Table 14.

Insights into performance evaluation of lightweight YOLO models

The evaluation of lightweight YOLO models is fundamentally about navigating the accuracy-efficiency trade-off. A key insight is that simply reducing model size or GFLOPs is insufficient; the goal is to achieve the best possible accuracy under given resource constraints. This often means that a model with slightly higher GFLOPs but significantly better mAP, or a model with slightly lower mAP but dramatically higher FPS on target hardware, might be preferred depending on the application.

• The interplay of metrics: Precision and recall are often inversely related, and their balance is captured by the F1-score or mAP. For lightweight models, the challenge is to maintain a high mAP (especially mAP50–95, indicating good localization) while drastically reducing GFLOPs and parameters, and achieving high FPS on edge devices. Researchers often employ techniques like pruning, quantization, and efficient architectural designs to optimize this complex interplay.

• Context-dependent prioritization: No single metric is universally superior. In autonomous driving, high recall for pedestrians is paramount, even if it means a slight drop in precision. In quality control, high precision might be prioritized to minimize false positives. For battery-powered IoT devices, energy consumption (often correlated with GFLOPs and parameters) becomes a primary concern, sometimes even over raw FPS.

• Beyond theoretical metrics: While GFLOPs and parameters provide a theoretical estimate of computational cost and model size, they do not directly translate to real-world performance. Actual FPS and latency are heavily influenced by hardware architecture, memory access patterns, software optimizations (e.g., TensorRT, OpenVINO), and batch size. Therefore, reporting actual inference speed on the target hardware is crucial.

Limitations of current evaluation practices and actionable recommendations for researchers

There are certain limitations in the performance evaluation of current lightweight YOLO models, mainly reflected in the allocation of indicator weights and scene adaptability.

• 1.

  Lack of standardized composite metrics: The existing research has not paid enough attention to the weight allocation of performance indicators and comprehensive evaluation methods. There isn’t a universally accepted scientific system to comprehensively reflect model performance under different application requirements, and the trade-offs between multiple indicators still need to be further explored.

  Actionable recommendation: Researchers should consider developing or adopting composite scores (e.g., a weighted sum of mAP, FPS, and energy consumption) tailored to specific application domains. Clearly state the rationale for chosen weights.

• 2.

  Scene-specific evaluation challenges: There are significant differences in the focus of performance indicators in different application scenarios. For example, in scenarios with high real-time requirements, inference speed (FPS/Latency) is a key indicator, while in scenarios with high accuracy requirements, mAP or F1-score is more important.

  Actionable recommendation: Always define the target application scenario and its primary constraints (e.g., “real-time detection on a Jetson Nano for pedestrian safety”). This context should explicitly guide the selection and prioritization of evaluation metrics.

• 3.

  Inconsistent benchmarking environments: Performance metrics like FPS are highly dependent on the specific hardware, software stack, and even input resolution. Inconsistent reporting makes direct comparison across articles challenging.

  Actionable Recommendation: When reporting efficiency metrics (FPS, latency, energy), always specify the exact hardware platform (e.g., “NVIDIA Jetson Nano, 2 GB, with TensorRT 8.x”), software environment, input resolution, and batch size. This ensures reproducibility and comparability.

• 4.

  Over-reliance on single metrics: Some studies might overemphasize one metric (e.g., GFLOPs) while neglecting others (e.g., actual FPS or mAP on challenging datasets).

  Actionable recommendation: Always present a suite of metrics covering both accuracy and efficiency. A table summarizing mAP, GFLOPs, Parameters, and FPS (on a specified device) provides a much clearer picture of the model’s overall performance.

Overall guidance for researchers on evaluation

To ensure robust and valuable research in lightweight YOLO models, researchers should:

Adopt a holistic view: Understand that lightweighting is not just about making models smaller, but about optimizing the entire system for a specific task and hardware.

Prioritize real-world relevance: Evaluate models not just on theoretical metrics, but on their actual performance in conditions mimicking the target deployment environment.

Be transparent and reproducible: Clearly document all experimental setups, including datasets, training parameters, hardware specifications, and software versions, to allow for independent verification of results.

Consider qualitative aspects: Beyond numerical metrics, assess the model’s robustness to noise, occlusions, varying lighting conditions, and its generalization ability to unseen data.

By adhering to these principles, researchers can contribute more meaningfully to the field, providing insights that are not only scientifically sound but also practically actionable for real-world AI deployment.

From the above summary and analysis, the key insights, current challenges, and actionable recommendations for each research question are summarized in Table 15 below, aiming to provide valuable guidance for future research in lightweight YOLO models.