A generalized model for seed internal quality detection based on terahertz imaging technology combined with image compressed sensing and improved-real ESRGAN

• Published in: Microchemical journal 2025-01, Vol.208, p.112410, Article 112410
• Main Authors: Jin-li, Yang, Bin, Li, A-kun, Yang, Zhao-xiang, Sun, Xia, Wan, Aiguo, Ouyang, Yan-de, Liu
• Format: Article
• Language: English
• Subjects: Compressed sensing; CS-IRE; Improved-real ESRGAN; Plumpness; Seed; Terahertz time-domain spectroscopic imaging technology
• ISSN: 0026-265X
• DOI: 10.1016/j.microc.2024.112410

[Display omitted] •Reduce seed hollowing and defects, enhancing market value and planting efficiency.•Propose CS-IRE model. Combine compressive sensing with improved Real ESRGAN for efficient and clear terahertz imaging.•Achieves 12.5% measurement ratio, saving 87.5% imaging time; Show significantly improves in PSNR, SSIM and visual quality.•Validated on various seeds (sunflower, black sunflower, pumpkin, et al.) with fullness errors under 5.56%.•The CS-IRE model eliminates need for species-specific models, boosting agricultural efficiency and assessment reliability. The problem of internal defects in seeds seriously affects their market value and planting efficiency. A terahertz imaging model (CS-IRE) combining compressed sensing (CS) with improved Real ESRGAN (IRE) is proposed in this study, to fulfil the need for non-destructive, fast, and accurate detection of internal quality of seeds in agriculture. It is also used for internal quality detection of a wide range of seeds. Firstly, terahertz imaging and Bernoulli random sampling are performed and high-speed acquisition of seed images is realized. The ADMM_TV algorithm is used for image reconstruction. Then the problem of the slow speed of traditional imaging is mitigated. Second, the problem of missing details in Real ESRGAN super-resolution reconstructed images is alleviated by utilizing the IRE model. The detail reproduction of the image is enhanced. The experimental results show that the CS-IRE model significantly reduces the imaging time at a measurement ratio of 12.5 %. The reconstructed images perform well in terms of PSNR, SSIM and visual quality. Further validation showed that the average error in fullness of the different seeds do not exceed 5.56 %. And the combined average error is only 3.98 %. These results demonstrate that the CS-IRE model significantly improves the detail restoration and structural similarity of images. It also has high detection accuracy in different structural feature seeds. In summary, the CS-IRE model proposed in this study combines terahertz imaging, compressed sensing and super-resolution imaging techniques. It provides a fast, nondestructive and accurate method for seed internal quality detection. It demonstrates its wide application prospects and practical value in the field of seed detection.