Photoelectric Target Detection Algorithm Based on NVIDIA Jeston Nano
<p>The data acquisition process of photoelectric target laser active imaging detection system.</p> "> Figure 2
<p>Sample images of a partial dataset. (<b>a</b>) Active image labeled with a true target in the dataset, (<b>b</b>) part of the true target dataset, and (<b>c</b>) part of the false target dataset.</p> "> Figure 3
<p>Algorithm flowchart of this article.</p> "> Figure 4
<p>Image differentiation. (<b>a</b>) Active image, (<b>b</b>) passive image, and (<b>c</b>) differential image.</p> "> Figure 5
<p>Three-dimensional view of the pixel values belonging to real and pseudo targets. (<b>a</b>) True target local pixel value and (<b>b</b>) partial pseudo-target local pixel value.</p> "> Figure 6
<p>Histogram statistics in outdoor environment: (<b>a</b>) differential image containing the photoelectric target, (<b>b</b>) histogram of the differential image, and (<b>c</b>) image containing the segmentation of the photoelectric target threshold.</p> "> Figure 7
<p>Photoelectric device echo shape.</p> "> Figure 8
<p>Comparison of before and after shape screening (<b>a</b>) image after threshold segmentation and (<b>b</b>) image after shape measurement.</p> "> Figure 9
<p>Comparison between the proposed and traditional methods: (<b>a1</b>,<b>a2</b>) active image, (<b>b1</b>,<b>b2</b>) passive image, (<b>c1</b>,<b>c2</b>) maximum entropy threshold segmentation, (<b>d1</b>,<b>d2</b>) iterative threshold segmentation, (<b>e1</b>,<b>e2</b>) Otsu-based threshold segmentation, (<b>f1</b>,<b>f2</b>) proposed threshold segmentation.</p> "> Figure 9 Cont.
<p>Comparison between the proposed and traditional methods: (<b>a1</b>,<b>a2</b>) active image, (<b>b1</b>,<b>b2</b>) passive image, (<b>c1</b>,<b>c2</b>) maximum entropy threshold segmentation, (<b>d1</b>,<b>d2</b>) iterative threshold segmentation, (<b>e1</b>,<b>e2</b>) Otsu-based threshold segmentation, (<b>f1</b>,<b>f2</b>) proposed threshold segmentation.</p> "> Figure 10
<p>Network structure of the teacher and student models. (<b>a</b>) Resnet18 network structure (teacher model) and (<b>b</b>) Shuffv2_x0_5 network structure (student model).</p> "> Figure 11
<p>Comparison of the number of candidate regions produced by different threshold segmentation methods: (<b>a</b>) active image, (<b>b</b>) passive image, (<b>c</b>) Otsu-based segmentation, (<b>d</b>) candidate area segmentation based on the Otsu threshold, (<b>e</b>) proposed threshold segmentation effect, (<b>f</b>) candidate area segmentation based on the proposed threshold.</p> "> Figure 12
<p>Forward reasoning comparison test: (<b>a</b>) active image, (<b>b</b>) inference result of ShuffNet, (<b>c</b>) inference result of ShuffEng.</p> "> Figure 13
<p>Comparison of running time between single-target area and multi-target area. (<b>A</b>) ShuNet inference for multi-target regions, (<b>B</b>) ShuNet inference for single-target regions, (<b>C</b>) ShuEng inference for multi-target regions, and (<b>D</b>) ShuEng inference for single-target regions.</p> ">
Abstract
:1. Introduction
2. Materials and Equipment
2.1. Data Acquisition Equipment and Processes
2.2. Dataset
2.3. Experimental Environment
3. Methods
3.1. Image Preprocessing
3.2. Candidate Region Discrimination
- The large difference in the number of layers between the teacher and the student models in knowledge distillation has a negative effect on the distillation results [20]. Since the hardware platform has small computing power, the student model will be a lightweight network, so the teacher model can only be a low-level network, such as ResNet18, ResNet34, and ResNet50.
- Since the image pixel of this training set is small and the target feature information is relatively single, increasing the number of convolutional layers may discard part of the acquired low-level feature information. The prediction accuracy of ResNet18, ResNet34, and ResNet50 after training is 99.54%, 99.17%, and 98.72%, respectively.
3.3. Model Acceleration
4. Results
4.1. Analysis of Threshold Segmentation Results
4.2. Analysis of Knowledge Distillation
4.3. Analysis of the Impact of Preprocessing on the Algorithm
4.4. TensorRT Acceleration and Algorithm Detection Results
5. Conclusions
Author Contributions
Funding
Informed Consent Statement
Conflicts of Interest
References
- Huayan, S.; Laixian, Z.; Yanzhong, Z.; Yonghui, Z. Progress of Free-Space Optical Communication Technology Based on Modulating Retro-Reflector. Laser Optoelectron. Prog. 2013, 50, 040004. [Google Scholar] [CrossRef]
- Laixian, Z.; Hua-yan, S.; Guihua, F.; Yanzhong, Z.; Yonghui, Z. Progress in free space optical communication technology based on cat-eye modulating retro-reflector. Chin. J. Opt. Appl. Opt. 2013, 6, 681–691. [Google Scholar]
- Mieremet, A.-L.; Schleijpen, R.-M.-A.; Pouchelle, P.-N. Modeling the detection of optical sights using retroreflection. Proc. SPIE 2008, 6950, 69500. [Google Scholar]
- Auclair, M.; Sheng, Y.; Fortin, J. Identification of Targeting Optical Systems By Multiwavelength Retroreflection. Opt. Eng. 2013, 52, 54301. [Google Scholar] [CrossRef]
- Anna, G.; Goudail, F.; Dolfi, D. General state contrast imaging: An optimized polarimetric imaging modality insensitive to spatial intensity fluctuations. Opt. Soc. Am. A 2012, 29, 892–900. [Google Scholar] [CrossRef] [PubMed]
- Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet classification with deep convolutional neural networks. In Proceedings of the Neural Information Processing Systems Conference, Lake Tahoe, NV, USA, 3–6 December 2012; pp. 1097–1105. [Google Scholar]
- Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv 2014, arXiv:1409.1556. [Google Scholar]
- He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the Neural Information Processing Systems Conference, Barcelona, Spain, 5–10 December 2016; pp. 770–778. [Google Scholar]
- Sun, D.; Wang, Y.; Ni, D.; Wang, T. AutoPath: Image-Specific Inference for 3D Segmentation. Front. Neurorobotics 2020, 14, 49. [Google Scholar] [CrossRef] [PubMed]
- Shao, H.; Wang, C.; Fu, Z.; Liu, Z. High-Resolution Distributed Radiation Detector System Assisted by Intelligent Image Recognition. Front. Phys. 2021, 9, 693838. [Google Scholar] [CrossRef]
- Ke, X. Research on Hidden Camera Detection and Recognization Method Based on Machine Vision; Huazhong University of Science and Technology: Wuhan, China, 2020. [Google Scholar]
- Liu, C.; Zhao, C.M.; Zhang, H.Y.; Zhang, Z.; Zhai, Y.; Zhang, Y. Design of an Active Laser Mini-Camera Detection System using CNN. IEEE Photonics J. 2019, 11, 1. [Google Scholar] [CrossRef]
- Huang, J.H.; Zhang, H.Y.; Wang, L.; Zhang, Z.; Zhao, C. Improved YOLOv3 Model for miniature camera detection. Opt. Laser Technol. 2021, 142, 107133. [Google Scholar] [CrossRef]
- Yanzhong, Z.; Huayan, S.; Xiaqiong, Y.; Mengshan, F. Three-Dimensional Analytical Formula for Oblique and Off-Axis Gaussian Beams Propagating through a Cat-Eye Optical Lens. Chin. Phys. Lett. 2010, 27, 034101. [Google Scholar] [CrossRef]
- Ma, Y.; Tang, J. Foundations of Deep Learning. In Deep Learning on Graphs; Cambridge University Press: Cambridge, UK, 2021; pp. 43–72. [Google Scholar] [CrossRef]
- Hinton, G.; Vinyals, O.; Dean, J. Distilling the Knowledge in a Neural Network. arXiv.org. Available online: https://arxiv.org/abs/1503.02531 (accessed on 10 August 2022).
- Lijun, Z.; Yu, L.; Lu, B.; Fei, L.; Yawei, W. Using TensorRT for deep learning and inference applications. J. Appl. Opt. 2020, 41, 337–341. [Google Scholar] [CrossRef]
- Shafi, O.; Rai, C.; Sen, R.; Ananthanarayanan, G. Demystifying TensorRT: Characterizing Neural Network Inference Engine on Nvidia Edge Devices. In Proceedings of the 2021 IEEE International Symposium on Workload Characterization (IISWC), Storrs, CT, USA, 7–9 November 2021; pp. 226–237. [Google Scholar] [CrossRef]
- Wang, X.; Jun, Z.; Wang, S. The Cat-Eye Effect Target Recognition Method Based on Visual Attention. Chin. J. Electron. 2019, 28, 1080–1086. [Google Scholar] [CrossRef]
- Romero, A.; Ballas, N.; Kahou, S.E.; Chassang, A.; Gatta, C.; Bengio, Y. Fitnets: Hints for thin deep nets. arXiv 2014, arXiv:1412.6550. [Google Scholar]
- Jeong, E.; Kim, J.; Tan, S.; Lee, J.; Ha, S. Deep Learning Inference Parallelization on Heterogeneous Processors with TensorRT. IEEE Embed. Syst. Lett. 2021, 14, 15–18. [Google Scholar] [CrossRef]
- Ningning, M.A.; Xiangyu, Z.; Haitao, Z.; Sun, J. ShuffleNet V2: Practical guidelines for efficient CNN architecture design. In Proceedings of the Computer Vision-ECCV 2018, Munich, Germany, 8–14 September 2018; pp. 116–131. [Google Scholar] [CrossRef]
- Iandola, F.N.; Song, H.; Moskewicz, M.W.; Ashraf, K.; Dally, W.J.; Keutzer, K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and<0.5 MB model size. arXiv 2022, arXiv:1602.07360. [Google Scholar]
- Han, K.; Wang, Y.; Tian, Q.; Guo, J.; Xu, C.; Xu, C. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, 14–19 June 2020; pp. 1580–1589. [Google Scholar]
- Yang, L.; Jiang, H.; Cai, R.; Wang, Y.; Song, S.; Huang, G.; Tian, Q. Condensenet v2: Sparse feature reactivation for deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, TN, USA, 20–25 June 2021; pp. 3569–3578. [Google Scholar]
- Jeong, E.; Kim, J.; Ha, S. TensorRT-based Framework and Optimization Methodology for Deep Learning Inference on Jetson Boards. Free Access 2022. [Google Scholar] [CrossRef]
Network | Top-1% | Parameter Quantity (M) |
---|---|---|
VGG16 | 99.94% | 134.27 |
AlexNet | 99.74% | 57.00 |
ResNet18 | 99.54% | 23.51 |
Network | Top-1% | KD- VGG16 Top-1% | KD- AlexNet Top-1% | KD- ResNet18 Top-1% | Parameter Quantity (M) | Calculated Amount (M) | Inference Time (s) |
---|---|---|---|---|---|---|---|
Shuffv2_x0_5 | 98.84% | 98.09% | 97.79% | 99.05% | 0.34 | 2.95 | 0.0783 |
Shuffv2_x1_0 | 99.26% | 98.69% | 98.84% | 99.63% | 1.26 | 11.62 | 0.0926 |
Shuffv2_x1_5 | 99.50% | 98.94% | 99.45% | 99.71% | 2.48 | 24.07 | 0.0986 |
Shuffv2_x2_0 | 99.67% | 99.44% | 99.59% | 99.73% | 5.35 | 47.62 | 0.1134 |
Squeezent1_0 | 97.86% | 88.34% | 63.18% | 97.94% | 0.73 | 41.74 | 0.0727 |
Squeezent1_1 | 91.20% | 96.03% | 81.81% | 95.23% | 0.72 | 16.05 | 0.0682 |
GhostNet | 97.94% | 99.19% | 99.04% | 98.99% | 3.90 | 14.26 | 0.0910 |
CondenseNetv2 | 95.12% | 98.54% | 96.73% | 98.69% | 7.26 | 169.0 | - |
Method | Number of Candidate Regions | Detection Rate (%) | False Alarm Rate (%) | Average Inference Time |
---|---|---|---|---|
Otsu | 1243 | 93.49% | 9.4% | 0.1781 |
Ours | 167 | 96.74% | 4.8% | 0.0822 |
Model | Detection Rate | False Alarm Rate |
---|---|---|
ShuffNet | 97.15% | 4.87% |
ShuffEng | 97.15% | 4.87% |
Single-Target Region Inference Time (/s) | Multi-Target Region Inference Time (/s) | |
---|---|---|
ShuffNet | 0.0788 | 0.1293 |
ShuffEng | 0.0341 | 0.0436 |
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |
© 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Zhang, S.; Zhang, L.; Sun, H.; Guo, H. Photoelectric Target Detection Algorithm Based on NVIDIA Jeston Nano. Sensors 2022, 22, 7053. https://doi.org/10.3390/s22187053
Zhang S, Zhang L, Sun H, Guo H. Photoelectric Target Detection Algorithm Based on NVIDIA Jeston Nano. Sensors. 2022; 22(18):7053. https://doi.org/10.3390/s22187053
Chicago/Turabian StyleZhang, Shicheng, Laixian Zhang, Huayan Sun, and Huichao Guo. 2022. "Photoelectric Target Detection Algorithm Based on NVIDIA Jeston Nano" Sensors 22, no. 18: 7053. https://doi.org/10.3390/s22187053
APA StyleZhang, S., Zhang, L., Sun, H., & Guo, H. (2022). Photoelectric Target Detection Algorithm Based on NVIDIA Jeston Nano. Sensors, 22(18), 7053. https://doi.org/10.3390/s22187053