Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization
Open accessThe DOI matches its Crossref record; 43 of the 43 references it cites were found in a registry or an index; 5 of the 8 sources its claims cite were checked for support.
Abstract
Most real-time computer vision applications heavily rely on Convolutional Neural Network (CNN) based models, for image classification and recognition. Due to the computationally and memory-intensive nature of the CNN model, it’s challenging to deploy on resource-constrained Internet of Things (IoT) devices to enable Edge intelligence for real-time decision-making. Edge intelligence requires minimum inference latency, memory footprint, and energy-efficient model. This work aims to develop an energy-efficient deep learning accelerator using a 3-stage pipeline: Training, Weight-pruning, and Quantization to reduce the model size and optimize the resources. First, we employ YOLOv3, a CNN architecture to detect objects in an image on the trained data. In addition, a sparse network of YOLO has been created by using pruning, which helps to improve the network’s performance and efficiency by reducing the computational requirements. Finally, we utilize 8-bit quantization to reduce the precision of the weights and activations, in a neural network. The evaluation of our proposed model shows that combining pruning and 8-bit quantization improves the efficiency and performance of the model. While pruning shows a decline of 80.39% in model parameters. The combination of 8-bit quantization results in an improvement in inference latency by 22.72% compared to existing SQuantization approach and a reduction of energy consumption by 29.41%.
Overview
CoChat Summary
TL;DR
The paper proposes a three-stage pipeline—training, pruning, and 8-bit k-means quantization—to compress a YOLOv3-SPP object detector for IoT edge deployment on COCO2017. It reports reduced parameters, lower inference latency, and 29.41% lower energy consumption while maintaining mAP@0.5 around 61.86%.
Key findings
- Pruning at an 80% ratio reduced parameters from 62,998,749 to 12,350,941 (abstract reports an 80.39% decline) and channels from 13,376 to 3,535/3,536 (actual pruning ratio 0.736), while mAP@0.5 increased from 61% to 61.85%; precision was 0.6970, recall 0.74, and F1 0.7178.
- Adding 8-bit k-means quantization yielded mAP 61.86% (61.8676%), inference latency of 17 ms, and a 29.41% reduction in energy consumption; the abstract also reports a 22.72% inference-latency improvement compared with SQuantization.
- The baseline YOLOv3-SPP on COCO2017 had inference speeds of 22 ms for 416×416 and 43 ms for 608×608; training validation loss decreased from 14.5 to 6.27 for 416×416 and from 12.7 to 4.95 for 608×608, with mAP increasing to 57.3% and 61%, respectively.
Methods
- Three-stage pipeline: initial training of pretrained YOLOv3-SPP on COCO2017; channel/weight pruning with sparsity 0.1 and pruning ratios up to 80%, performed without retraining; then 8-bit integer uniform quantization using k-means clustering on pruned weights.
- Data and setup: COCO2017 (118k train, 40.7k test, 5k validation; 80 object categories), input sizes 416×416 and 608×608, baseline batch size 16 and 100 epochs; quantization algorithm uses bit width 8, batch size 64, and 10 epochs. Hardware included Nvidia V100 Tensor Core GPUs and an Intel i5 8th-generation CPU; implementation used PyTorch, NumPy, Matplotlib, and scikit-learn.
- Main analysis: mAP@0.5, precision, recall, F1, parameter count, inference latency, and energy consumption, with comparison against SQuantization.
Limitations
- The paper does not state a dedicated limitations section. It explicitly notes that pruning is performed without retraining, which may constrain accuracy recovery.
- Evident from the text: evaluation is limited to COCO2017 and YOLOv3-SPP, and training/testing used Nvidia V100 and Intel i5 hardware rather than resource-constrained IoT/edge devices, so edge-deployment claims are not directly validated on such hardware.
- Evident from the text: some reported numbers are inconsistent—e.g., pruning ratio described as 80% but actual 73.6%, channels 3,535 vs 3,536, and Table 9 before-pruning mAP 0.702388 vs text 61%—which complicates interpretation.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
3 Valid · 9 Not valid · 3 Not checked · 46 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“4 bit SQuantizer gives 50.3 mAP compared to baseline mAP 52.8 with slight lesser than the orginal mAP value.”
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > 4.3.5 Comparative analysis
What the sources say
- Disputedcites [40] Park MS, Xu X, Brick C (2018) Squantizer: Simultaneous learning for both sparse and low-precision neural networks. arXiv preprint arXiv:1812.08301 sourceExperimental — accuracy not yet measured
Table 10 (section 4.2, p. 8) reports a baseline of 52.8 mAP, but assigns 50.3 mAP to the 2-bit model. The 4-bit model achieves 52.0 mAP instead. (LLM, conf=0.95)
Checked by a reasoning model — may be wrong.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“It is evident that floating point add process takes three orders of magnitudes of add operation [19].”
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > 1 Introduction
What the sources say
- Disputedcites [19] Han S, Pool J, Tran J, Dally W (2015) Learning both weights and connections for efficient neural network. Adv Neural Inf Process Syst 28 sourceExperimental — accuracy not yet measured
Our copy's energy table (Table 1 / Figure 1, 1 Introduction, p. 2) gives 32 bit int ADD = 0.1 pJ (relative cost 1) and 32 bit float ADD = 0.9 pJ (relative cost 9), i.e. a floating-point add costs about 9x — one order of magnitude — more than an integer add, not three orders. The only '3 orders of magnitude' statement in the paper is the Figure 1 caption (1 Introduction, p. 1): 'Memory access is 3 orders of magnitude more energy expensive than simple arithmetic,' which concerns off-chip DRAM access (640 pJ) versus arithmetic, and is itself credited to reference [7]. So the source contradicts the magnitude attributed to it for floating-point addition. (LLM, conf=0.80)
Checked by a reasoning model — may be wrong.
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“with 61.85% of mAP@0.5”
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > 5 Conclusion
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“inference latency of 17ms”
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > 5 Conclusion
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“As we can see from Figure5 b), validation loss started decrementing from 12.7 to 4.95 and mean average precision gradually increased from 30.3% to 61%.”
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > 5 Conclusion
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
32 more claims
Create a free account to see all 61 claims and every source verdict.
- Probably not a substantial claim. 15
- Borderline: one of the tests for a substantial claim is unclear. 2
- It may state two separate claims as one. 12
- The quote only partly states this claim. 3
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > 4.3.5 Comparative analysis
“Table 9 KPIs: mAP, Inference /prediction time, parameters before & after pruning Metric Before After mAP 0.702388 0.7 Parameters 62998749 12350941 Inference 0.2401 0.1483”
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > Acceleration Ratio =
“Table 8 Pruning percentage and corresponding mAP Pruning (%) mAP@0.5 Precision Recall F1 0 0.61 0.675 0.73 0.701423 70 0.608 0.669 0.745 0.7049 75 0.6156 0.6598 0.769 0.7102 80 0.6185 0.6970 0.74 0.7178”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > 4.3.3 Pruning the model
“Total number of channels reduced from 13376 to 3535 with the pruning ratio of 73.6%.”
§ Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization > Acceleration Ratio =
“Number of channels reduced from 13376 to 3536 with the actual pruning ration of 0.736.”
Citations
Cites
Han, S., Mao, H., & Dally, W. J. (2015) · arXiv · doi:10.48550/arxiv.1510.00149 · Cited by 4,662
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference
Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., & Kalenichenko, D. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00286 · Cited by 4,114
Convergence of Edge Computing and Deep Learning: A Comprehensive Survey
Wang, X., Han, Y., Leung, V. C. M., Niyato, D., Yan, X., & Chen, X. (2020) · IEEE Communications Surveys & Tutorials · doi:10.1109/comst.2020.2970550 · Cited by 1,498 · cited 1×
“Edge computing [1] refers processing of data at the edge of a network rather than sending it to a central location for processing.”
Learning both Weights and Connections for Efficient Neural Networks
Han, S., Pool, J., Tran, J., & Dally, W. J. (2015) · arXiv · doi:10.48550/arxiv.1506.02626 · Cited by 1,481
Pruning Convolutional Neural Networks for Resource Efficient Inference
Molchanov, P. A., Tyree, S., Karras, T., Aila, T., & Kautz, J. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1611.06440 · Cited by 1,239
Filter Pruning via Geometric Median for Deep Convolutional Neural Networks Acceleration
He, Y., Liu, P., Wang, Z., Hu, Z., & Yang, Y. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2019.00447 · Cited by 1,217
Edge Intelligence: The Confluence of Edge Computing and Artificial Intelligence
Deng, S., Zhao, H., Fang, W., Yin, J., Dustdar, S., & Zomaya, A. Y. (2020) · IEEE Internet of Things Journal · doi:10.1109/jiot.2020.2984887 · Cited by 1,147
Pruning and quantization for deep neural network acceleration: A survey
Liang, T., Glossner, J., Wang, L., Shi, S., & Zhang, X. (2021) · Neurocomputing · doi:10.1016/j.neucom.2021.07.045 · Cited by 943
Designing Energy-Efficient Convolutional Neural Networks Using Energy-Aware Pruning
Yang, T.-J., Chen, Y.-H., & Sze, V. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.643 · Cited by 732
DeepTrust^RT: Confidential Deep Neural Inference Meets Real-Time!
Babar, M. F., & Hasan, M. (2024) · DROPS (Schloss Dagstuhl – Leibniz Center for Informatics) · doi:10.4230/lipics.ecrts.2024.13 · Cited by 708
Research works printed but not linked: 2
- Maini DS, Aggarwal AK (2018) Camera position estimation using 2D image dataset. Int J Innov Eng Technol 10(2):199-203. http://dx.doi.org/10.21172/ijiet.102.29 · doi:10.21172/ijiet.102.29
- Aggarwal AK (2015) A hybrid approach to GPS improvement in urban canyons? Int J Eng Sci Res Technol 4(10):358–363
Cited by
Kabir, R., Watanobe, Y., Ding, D., Islam, M. R., & Naruse, K. (2025) · IEEE Access · doi:10.1109/access.2025.3619776 · Cited by 24
Saheed, Y. K., & Chukwuere, J. E. (2025) · IEEE Internet of Things Journal · doi:10.1109/jiot.2025.3648649 · Cited by 29
Deng, W., Qin, Q., Zhao, J., Yu, Y., Huang, Y., Dong, H., Cao, F., & Li, Z. (2026) · Food Chemistry · doi:10.1016/j.foodchem.2026.147961 · Cited by 15
A systematic literature review of lightweight YOLO models for object detection
Wang, L., Wang, H., Letchmunan, S., Xiao, R., Ahmed, O. H., & Liu, Z. (2025) · PeerJ Computer Science · doi:10.7717/peerj-cs.3357 · Cited by 18
Rivera, E. F., Guerrero-Vásquez, L. F., & Robles-Bykbaev, V. (2026) · Technologies · doi:10.3390/technologies14010076 · Cited by 18
Metadata
- Authors
- Naveen, S., & Kounte, M. R.
- Affiliations
- Manipal Academy of Higher Education (IN)
- REVA University (IN)
- Corresponding author
- Naveen, Soumyalatha
- Venue
- Multimedia Tools and Applications
- Published
- 2024-12-26
- Volume
- 84
- Issue
- 9
- Pages
- 5435-5455
- ISSN
- 1380-7501 (ISSN-L)
- 1573-7721
- OpenAlex
- W4405783794
- Semantic Scholar
- 46fb2055ca4a2cb77fd4518e794c3cf36238d8a4
- Cited by
- 33
- References
- 40
- Citations by year
- 2026: 21 · 2025: 12
- FWCI
- 4.61
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- Multimedia Tools and Applications (publishedVersion, cc-by)
- Funders
- Manipal Academy of Higher Education, Manipal
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Keywords
- convolutional neural networks, edge AI, weight pruning, 8-bit quantization, YOLOv3, object detection, energy-efficient inference, inference latency, model compression, sparse neural networks
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization. https://app.cochat.ai/papers/10.1007/s11042-024-20523-1
- The paper
- Naveen, S., & Kounte, M. R. (2024). Optimized Convolutional Neural Network at the IoT edge for image detection using pruning and quantization. Multimedia Tools and Applications. https://doi.org/10.1007/s11042-024-20523-1
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.