{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,3]],"date-time":"2026-09-03T11:12:40Z","timestamp":1788433960614,"version":"build-2803163510"},"reference-count":37,"publisher":"MDPI AG","issue":"17","license":[{"start":{"date-parts":[[2026,8,31]],"date-time":"2026-08-31T00:00:00Z","timestamp":1788134400000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"name":"USDA\u2014Agricultural Research Service (ARS) through a research support agreement between the USDA\u2014ARS and the University of Georgia","award":["58-6040-4-033"],"award-info":[{"award-number":["58-6040-4-033"]}]}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["Sensors"],"abstract":"<jats:p>Low-precision floating-point computation has become central to efficient artificial intelligence, yet its behavior for compact transformer-based hyperspectral imaging (HSI) models remains underexplored. In this work, we present a controlled comparative study of fully quantized training (FQT) and post-training quantization (PTQ) for pixel-wise foreign plastic object (FPO) classification in poultry hyperspectral data. Using a fixed state-of-the-art spatial\u2013spectral transformer backbone, a common mixed-precision strategy, and identical training and inference protocols, we evaluate FP32, FP16, BF16, FP8, and NVFP4 across predictive performance, model compression, training efficiency, and inference efficiency. The results show that mixed-precision FQT remains highly robust across the tested precision spectrum, with all reduced-precision configurations staying within 0.63 percentage points of the FP32 baseline in overall accuracy while consistently outperforming PTQ at matched precisions. Across the evaluated formats, BF16 provides the closest accuracy to FP32, whereas FP8 offers a particularly favorable balance between accuracy preservation and reduced precision, while model compression increases progressively to 3.69\u00d7 under NVFP4. The computational benefits, however, are strongly workload dependent. Native FP8\/FP4 hardware support does not automatically improve training throughput for this compact model at moderate workloads, and the larger training batches required to better utilize low-precision hardware can degrade predictive performance. In contrast, large-batch inference can effectively exploit FP8 and NVFP4 without affecting predictive accuracy. An ablation study further shows that selective retention of numerically sensitive modules in FP32 is essential for stable ultra-low-precision operation. Overall, the findings demonstrate that low precision is a viable but workload-dependent design choice for compact HSI transformers, with FQT providing greater accuracy robustness than PTQ and FP8, offering a favorable overall accuracy\u2013efficiency trade-off.<\/jats:p>","DOI":"10.3390\/s26175531","type":"journal-article","created":{"date-parts":[[2026,9,1]],"date-time":"2026-09-01T03:19:26Z","timestamp":1788232766000},"page":"5531","update-policy":"https:\/\/doi.org\/10.3390\/mdpi_crossmark_policy","source":"Crossref","is-referenced-by-count":0,"title":["Fully Quantized Training vs. Post-Training Quantization for a Small Hyperspectral Transformer Model for Pixel-Level Foreign Plastic Object Classification"],"prefix":"10.3390","volume":"26","author":[{"ORCID":"https:\/\/orcid.org\/0009-0009-3668-3441","authenticated-orcid":false,"given":"Zirak","family":"Khan","sequence":"first","affiliation":[{"name":"School of Computing, University of Georgia, Athens, GA 30602, USA"}],"role":[{"vocabulary":"crossref","role":"author"},{"role":"conceptualization","vocabulary":"credit"},{"role":"methodology","vocabulary":"credit"},{"role":"software","vocabulary":"credit"},{"role":"writing-original-draft","vocabulary":"credit"},{"role":"writing-review-editing","vocabulary":"credit"}]},{"ORCID":"https:\/\/orcid.org\/0000-0001-8143-6559","authenticated-orcid":false,"given":"Seung-Chul","family":"Yoon","sequence":"additional","affiliation":[{"name":"Quality and Safety Research Unit, U.S. National Poultry Research Center, U.S. Department of Agriculture\u2014Agricultural Research Service, Athens, GA 30605, USA"}],"role":[{"vocabulary":"crossref","role":"author"},{"role":"conceptualization","vocabulary":"credit"},{"role":"data-curation","vocabulary":"credit"},{"role":"writing-review-editing","vocabulary":"credit"},{"role":"project-administration","vocabulary":"credit"},{"role":"funding-acquisition","vocabulary":"credit"}]},{"given":"Suchendra M.","family":"Bhandarkar","sequence":"additional","affiliation":[{"name":"School of Computing, University of Georgia, Athens, GA 30602, USA"}],"role":[{"vocabulary":"crossref","role":"author"},{"role":"conceptualization","vocabulary":"credit"},{"role":"project-administration","vocabulary":"credit"},{"role":"funding-acquisition","vocabulary":"credit"},{"role":"writing-review-editing","vocabulary":"credit"}]}],"member":"1968","published-online":{"date-parts":[[2026,8,31]]},"reference":[{"key":"ref_1","doi-asserted-by":"crossref","first-page":"217","DOI":"10.1109\/TGRS.2012.2201730","article-title":"Hyperspectral Image Classification via Kernel Sparse Representation","volume":"51","author":"Chen","year":"2013","journal-title":"IEEE Trans. Geosci. Remote Sens."},{"key":"ref_2","doi-asserted-by":"crossref","unstructured":"Sandak, J., Sandak, A., Legan, L., Retko, K., Kav\u010di\u010d, M., Kosel, J., Poohphajai, F., Diaz, R.H., Ponnuchamy, V., and Sajin\u010di\u010d, N. (2021). Nondestructive Evaluation of Heritage Object Coatings with Four Hyperspectral Imaging Systems. Coatings, 11.","DOI":"10.3390\/coatings11020244"},{"key":"ref_3","doi-asserted-by":"crossref","unstructured":"Chung, S., and Yoon, S.C. (2021). Detection of Foreign Materials on Broiler Breast Meat Using a Fusion of Visible Near-Infrared and Short-Wave Infrared Hyperspectral Imaging. Appl. Sci., 11.","DOI":"10.3390\/app112411987"},{"key":"ref_4","doi-asserted-by":"crossref","first-page":"5518615","DOI":"10.1109\/TGRS.2021.3130716","article-title":"SpectralFormer: Rethinking Hyperspectral Image Classification with Transformers","volume":"60","author":"Hong","year":"2022","journal-title":"IEEE Trans. Geosci. Remote Sens."},{"key":"ref_5","doi-asserted-by":"crossref","first-page":"1910","DOI":"10.1109\/JSTARS.2025.3643365","article-title":"Spatial-Spectral Transformer with Patch-Local Mixed-Axis 2-D Rotary Position Embedding for Hyperspectral Image Classification","volume":"19","author":"Khan","year":"2026","journal-title":"IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens."},{"key":"ref_6","doi-asserted-by":"crossref","unstructured":"Kwon, J.S., Lee, J.P., and Kim, W.Y. (2008, January 14\u201316). Real-time Detection of Foreign Objects Using X-Ray Imaging for Dry Food Manufacturing Line. Proceedings of the 2008 IEEE International Symposium on Consumer Electronics (ISCE 2008), Vilamoura, Portugal.","DOI":"10.1109\/ISCE.2008.4559552"},{"key":"ref_7","unstructured":"Ristau, K. (2022). Using Foreign Material Detection Methods to Assist the Food Industry. [Master\u2019s Thesis, Kansas State University]."},{"key":"ref_8","doi-asserted-by":"crossref","first-page":"19574","DOI":"10.1016\/j.heliyon.2023.e19574","article-title":"Detection and Prevention of Foreign Material in Food: A Review","volume":"9","author":"Payne","year":"2023","journal-title":"Heliyon"},{"key":"ref_9","doi-asserted-by":"crossref","first-page":"113947","DOI":"10.1016\/j.engappai.2026.113947","article-title":"Transformers Meet Hyperspectral Imaging: A Comprehensive Study of Models, Challenges and Open Problems","volume":"169","author":"Zhang","year":"2026","journal-title":"Eng. Appl. Artif. Intell."},{"key":"ref_10","doi-asserted-by":"crossref","first-page":"159","DOI":"10.1016\/j.compag.2011.09.008","article-title":"Line-scan hyperspectral imaging system for real-time inspection of poultry carcasses with fecal material and ingesta","volume":"79","author":"Yoon","year":"2011","journal-title":"Comput. Electron. Agric."},{"key":"ref_11","doi-asserted-by":"crossref","unstructured":"Horowitz, M. (2014, January 9\u201313). Computing\u2019s Energy Problem (and What We Can Do About It). Proceedings of the 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), San Francisco, CA, USA.","DOI":"10.1109\/ISSCC.2014.6757323"},{"key":"ref_12","doi-asserted-by":"crossref","unstructured":"Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., and Kalenichenko, D. (2018, January 18\u201323). Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference. Proceedings of the 2018 IEEE\/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Salt Lake City, UT, USA.","DOI":"10.1109\/CVPR.2018.00286"},{"key":"ref_13","doi-asserted-by":"crossref","unstructured":"Thiruvathukal, G.K., Lu, Y.H., Kim, J., Chen, Y., and Chen, B. (2022). A Survey of Quantization Methods for Efficient Neural Network Inference. Low-Power Computer Vision: Improve the Efficiency of Artificial Intelligence, Chapman and Hall\/CRC. [1st ed.].","DOI":"10.1201\/9781003162810"},{"key":"ref_14","unstructured":"Kalamkar, D., Mudigere, D., Mellempudi, N., Das, D., Banerjee, K., Avancha, S., Vooturi, D.T., Jammalamadaka, N., Huang, J., and Yuen, H. (2019). A Study of BFLOAT16 for Deep Learning Training. arXiv."},{"key":"ref_15","doi-asserted-by":"crossref","first-page":"1302","DOI":"10.1109\/TETC.2022.3187770","article-title":"A BF16 FMA Is All You Need for DNN Training","volume":"10","author":"Osorio","year":"2022","journal-title":"IEEE Trans. Emerg. Top. Comput."},{"key":"ref_16","unstructured":"Fishman, M., Chmiel, B., Banner, R., and Soudry, D. (2025, January 24\u201328). Scaling FP8 Training to Trillion-Token LLMs. Proceedings of the Thirteenth International Conference on Learning Representations (ICLR 2025), Singapore."},{"key":"ref_17","doi-asserted-by":"crossref","unstructured":"Hern\u00e1ndez-Cano, A., Garbaya, D., Schlag, I., and Jaggi, M. (2025, January 2\u20137). Towards Fully FP8 GEMM LLM Training at Scale. Proceedings of the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025), San Diego, CA, USA.","DOI":"10.52202\/085713-1896"},{"key":"ref_18","unstructured":"Peng, H., Wu, K., Wei, Y., Zhao, G., Yang, Y., Liu, Z., Xiong, Y., Yang, Z., Ni, B., and Hu, J. (2023). FP8-LM: Training FP8 Large Language Models. arXiv."},{"key":"ref_19","unstructured":"Micikevicius, P., Stosic, D., Burgess, N., Cornea, M., Dubey, P., Grisenthwaite, R., Ha, S., Heinecke, A., Judd, P., and Kamalu, J. (2022). FP8 Formats for Deep Learning. arXiv."},{"key":"ref_20","unstructured":"NVIDIA Corporation (2025). Introducing NVFP4 for Efficient and Accurate Low-Precision Inference, NVIDIA. Available online: https:\/\/developer.nvidia.com\/blog\/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference."},{"key":"ref_21","doi-asserted-by":"crossref","unstructured":"Chmiel, B., Fishman, M., Banner, R., and Soudry, D. (2025, January 2\u20137). FP4 All the Way: Fully Quantized Training of Large Language Models. Proceedings of the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025), San Diego, CA, USA.","DOI":"10.52202\/085713-3049"},{"key":"ref_22","first-page":"62937","article-title":"Optimizing Large Language Model Training Using FP4 Quantization","volume":"Volume 267","author":"Wang","year":"2025","journal-title":"Proceedings of the 42nd International Conference on Machine Learning, Vancouver, BC, Canada, 13\u201319 July 2025"},{"key":"ref_23","first-page":"4466","article-title":"Accurate Post Training Quantization with Small Calibration Sets","volume":"Volume 139","author":"Hubara","year":"2021","journal-title":"Proceedings of the 38th International Conference on Machine Learning, Virtual, 18\u201324 July 2021"},{"key":"ref_24","unstructured":"Singh, J. (2024). Empirical Evaluation of Edge AI Deployment Strategies Involving Black-Box and White-Box Operators. [Master\u2019s Thesis, Queen\u2019s University]."},{"key":"ref_25","unstructured":"Bengio, Y., L\u00e9onard, N., and Courville, A. (2013). Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation. arXiv."},{"key":"ref_26","doi-asserted-by":"crossref","first-page":"107856","DOI":"10.1016\/j.neunet.2025.107856","article-title":"A Survey of Low-Bit Large Language Models: Basics, Systems, and Algorithms","volume":"192","author":"Gong","year":"2025","journal-title":"Neural Netw."},{"key":"ref_27","doi-asserted-by":"crossref","unstructured":"Nagel, M., van Baalen, M., Blankevoort, T., and Welling, M. (November, January 27). Data-Free Quantization Through Weight Equalization and Bias Correction. Proceedings of the IEEE\/CVF International Conference on Computer Vision (ICCV 2019), Seoul, Republic of Korea.","DOI":"10.1109\/ICCV.2019.00141"},{"key":"ref_28","unstructured":"Banner, R., Nahshan, Y., and Soudry, D. (2019, January 8\u201314). Post Training 4-Bit Quantization of Convolutional Networks for Rapid-Deployment. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS 2019), Vancouver, BC, Canada."},{"key":"ref_29","unstructured":"Frantar, E., Ashkboos, S., Hoefler, T., and Alistarh, D. (2022). GPTQ: Accurate Post-Training Quantization for Generative Pre-Trained Transformers. arXiv."},{"key":"ref_30","unstructured":"Lin, J., Tang, J., Tang, H., Yang, S., Chen, W.M., Wang, W.C., Xiao, G., Dang, X., Gan, C., and Han, S. (2024, January 13\u201316). AWQ: Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration. Proceedings of the Machine Learning and Systems, Santa Clara, CA, USA."},{"key":"ref_31","unstructured":"Perez, S.P., Zhang, Y., Briggs, J., Blake, C., Levy-Kramer, J., Balanca, P., Luschi, C., Barlow, S., and Fitzgibbon, A.W. (2023, January 16). Training and Inference of Large Language Models Using 8-Bit Floating Point. Proceedings of the Workshop on Advancing Neural Network Training: Computational Efficiency, Scalability, and Resource Optimization (WANT@NeurIPS 2023), New Orleans, LA, USA."},{"key":"ref_32","unstructured":"NVIDIA Corporation (2026, March 23). Per-Tensor and Per-Block Scaling Strategies for Effective FP8 Training. NVIDIA Technical Blog. Available online: https:\/\/developer.nvidia.com\/blog\/per-tensor-and-per-block-scaling-strategies-for-effective-fp8-training\/."},{"key":"ref_33","doi-asserted-by":"crossref","unstructured":"Kimhi, M., Rozen, T., Mendelson, A., and Baskin, C. (2024). AMED: Automatic Mixed-Precision Quantization for Edge Devices. Mathematics, 12.","DOI":"10.3390\/math12121810"},{"key":"ref_34","unstructured":"(1985). IEEE Standard 754 for Binary Floating-Point Arithmetic (Standard No. IEEE Standard 754-1985)."},{"key":"ref_35","unstructured":"Rouhani, B.D., Zhao, R., More, A., Hall, M., Khodamoradi, A., Deng, S., Choudhary, D., Cornea, M., Dellinger, E., and Denolf, K. (2023). Microscaling Data Formats for Deep Learning. arXiv."},{"key":"ref_36","doi-asserted-by":"crossref","unstructured":"Dettmers, T., Lewis, M., Belkada, Y., and Zettlemoyer, L. (December, January 28). LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale. Proceedings of the Advances in Neural Information Processing Systems 35 (NeurIPS 2022), New Orleans, LA, USA.","DOI":"10.52202\/068431-2198"},{"key":"ref_37","doi-asserted-by":"crossref","unstructured":"Khan, Z., Yoon, S.C., and Bhandarkar, S.M. (2026). Parameter- and Compute-Efficient Spatial\u2013Spectral Transformer Framework for Pixel-Level Classification of Foreign Plastic Objects on Broiler Meat Using NIR\u2013Hyperspectral Imaging. Sensors, 26.","DOI":"10.3390\/s26082459"}],"container-title":["Sensors"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/www.mdpi.com\/1424-8220\/26\/17\/5531\/pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,9,3]],"date-time":"2026-09-03T10:15:22Z","timestamp":1788430522000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.mdpi.com\/1424-8220\/26\/17\/5531"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,8,31]]},"references-count":37,"journal-issue":{"issue":"17","published-online":{"date-parts":[[2026,9]]}},"alternative-id":["s26175531"],"URL":"https:\/\/doi.org\/10.3390\/s26175531","relation":{},"ISSN":["1424-8220"],"issn-type":[{"value":"1424-8220","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,8,31]]}}}