Non-destructive And Rapid Detection Of Fish Freshness Using Hyperspectral Imaging Technique

Freshness is an important index for the evaluation of fish quality and safety. Therefore, detection and evaluation of fish freshness is a critical issue for food safety control and the consumers’ health. The traditional freshness detection methods and techniques are usually time-consuming, tedious, destructive and not suitable for modern fish industry development. Looking for rapid, non-destructive, on-line and objective methods and techniques is very urgent. This dissertation is aimed to investigate the feasibility of hyperspectral imaging technique（400-1000 nm） coupled with chemometrics analysis and image processing algorithms for the detection and evaluation of grass carp fish fillet freshness quality. The main research contents and results are as follows.The use of hyperspectral imaging in tandem with classification algorithms and spectral pre-processing methods for discrimination of fresh samples, cold storage samples at 4 o C, and frozen samples at-20 o C and-40 o C was investigated. SIMCA, LS-SVM and PNN with the first derivative method showed good performances and the CCR value was improved from 88.57% to 94.29%. The simplified LS-SVM coupled with the first derivative method also displayed good classification capability. The CCR value was enhanced from 82.86% to 91.43%.The sensory evaluation scores were predicted by using hyperspectral imaging with LS-SVM model. The constructed LS-SVM model based on five optimal wavelengths（441 nm, 560 nm, 598 nm, 639 nm, and 684 nm） selected by SPA and thirteen image texture parameters extracted by GLGCM algorithm showed the best prediction performance with RPD of 4.230, R2 P of 0.944 and RMSEP of 0.703.The color（L^* and a^*） and firmness changes of grass carp fish fillets during cold and frozen stroage were measured using hyperspectral imaging with PLSR and LS-SVM algorithms. Based on full wavelengths analysis, both of PLSR and LS-SVM models showed good performances for the prediction of L^*（R2P = 0.906, RMSEP = 2.459; R2 P = 0.916, RMSEP = 2.876） and a^*（R2P = 0.887, RMSEP = 2.232; R2 P = 0.905, RMSEP = 2.253）. After MSC pre-processing, the constructed MSC-LS-SVM model exhibited better performance for the prediction of firmness of fish suffering from freezing-thawing cycles with R2 P of 0.932 and RMSEP of 1.351 N. The simplified SPA-LS-SVM model also showed better prediction ability for L^* and a^* with R2 P of 0.912 and 0.891, respectively. As to the firmness prediction, the simplified GA-LS-SVM model also displayed satisfactory results with R2 P of 0.941 and RMSEP of 1.229 N.The chemical information of TVB-N, TBA and K value was visualized with pseudo color using hyperspectral imaging and image processing algorithms. Nine feature wavelengths（420 nm, 466 nm, 523 nm, 552 nm, 595 nm, 615 nm, 717 nm, 850 nm and 955 nm） for indicating the changes of TVB-N in fish fillets were obtained by SPA. The simplified SPA-LS-SVM model showed better predictive power（R2P = 0.902 and RMSEP = 2.782 mg N/100 g）. Ten key wavelengths（444 nm, 475 nm, 553 nm, 577 nm, 590 nm, 623 nm, 710 nm, 795 nm, 847 nm and 937 nm） for representing the lipid oxidation in fish fillets were selected by RC method based on P LSR analysis. The simplified RC-MLR model also showed admirable capability with R2 P of 0.840 and RMSEP of 0.115 mg/kg. Seven important wavelengths（432 nm, 455 nm, 588 nm, 635 nm, 750 nm, 840 nm and 970 nm） for signifying K value changes in fish fillets were located by SPA. The simplified SPA-PLSR model exhibited satisfactory performance（R2P = 0.935 and RMSEP = 5.170%）. Six key wavelengths（435 nm, 565 nm, 660 nm, 815 nm, 870 nm and 970 nm） were selected by GA method for simultaneous determination of TVB-N, TBA and K value in fish fillets. Both of the constructed LS-SVM and MLR models showed excellent performances for synchronous measurement of TVB-N and K value with R2 more than 0.900 and RPD more than 3.000. The simplified better models were used to visualize the chemical values distribution in fish fillets by using image processing analysis. The dynamic variations of these chemical indicators were finally visualized and obtained.The microbial spoilage of fish fillets during cold storage was also evaluated using hyperspectral imaging technqiue based on the determination of TVC value and E. coli loads. With the increase of the TVC value and E. coli loads, the longitudinal shift of spectral reflectance value was improved. The LS-SVM model using full wavelengths showed better results with R2 P of 0.931, RPD of 3.891, RMSEP of 0.485 log10CFU/g and SWS of 0.360, respectively. The simplified SPA-LS-SVM and SPA-PLSR model showed similar prediction performance due to the same values of SWS. Six key wavelengths selected by RC method for the representation of E. coli loads were 424 nm, 451 nm, 545 nm, 567 nm, 585 nm and 610 nm. The simplified RC-MLR showed more superior performance for the prediction of E. coli loads with R₂ P of 0.870, RPD of 5.220 and RMSEP of 0.274 log₁₀ CFU/g.Based on these investigations, it h as been confirmed that hyperspectral imaging system coupled with chemometrics analysis and image processing method has the potential for rapid and non-destructive monitoring of fish freshness.