Rapid And Non-destructive Detection Of Marine Fish Quality Based On Spectroscopy And Hvperspectral Imaging Technique

Fish has always been regarded as a kind of popular food because it is easily-digestible and has high nutritional value. Fish is succulent with high moisture content, and spoilage occers easily in fish due to the activity of catabolic enzymes and microorganisms. Therefore the detection of fish quality and safety is significant. Traditional physical and chemical analytical methods are time-consuming, lengthy, destructive, and producing hazardous pollutant, thus they can no longer meet the requirements of rapid, non-destructive and real-time measurement for modern digital fishery development. This dissertation focuses on the application of spectroscopy and hyperspectral imaging technique, combined with chemometrics, image processing methods and data mining technologies to inspect the quality and safety of marine fish, providing technical support for digital fishery. Four parts are included:the discrimination of fresh and frozen-thawed turbot, the detection and visualization of the chilling storage time for turbot, the fast determination of Texture Profile Analysis (TPA) properties and water holding capacility for turbot, the mapping of fat and moisture contents in Atlantic salmon. The main research contents and results are as follows:(1) Fresh and frozen-thawed turbot samples were discriminated using visible and short-wave near-infrared hyperspectral imaging. Frozen treatments cover the different freezing temperature (fast frozen-70℃and slow frozen-20℃) and the different thawing times (frozen-thawed once and twice). Mean spectra were extracted from the hyperspectral images to be preprocessed by different pretreatment methods, then a novel method called competitive adaptive reweighted sampling (CARS) was employed to select the effective wavelengths from the full-spectrum. Meanwhile, principal component analysis (PCA) was conducted on the hyperspectral images, and textural variables were extracted from the first three principal component (PC) images. Finally, the linear method of partial least squares discriminant analysis (PLSDA), non-linear methods of back-propagation artificial neural network (BP-ANN) and least-square support vector machine (LS-SVM) were applied for discrimination based on the effective wavelengths, textural variables, combined effective wavelengths and textural variables, respectively. For the differentiation of fresh, fast frozen-thawed, slow frozen-thawed fish, the best overall correct classification rate was94.44%; for the differentiation of fresh, frozen-thawed once, frozen-thawed twice fish, the best overall correct classification rate was100%. (2) The detection and visualization of the chilling storage time for turbot was achieved by visible and short-wave near-infrared hyperspectral imaging. Mean spectra were extracted from the hyperspectral images to be correlated with the chilling storage time using partial least squares regression (PLSR), BP-ANN and LS-SVM. Good results were obtained by the linear PLSR model with correlation coefficient (Rp) of0.9849and root mean square error of prediction (RMSEP) of0.6799. Then the chilling storage time of each pixel in the hyperspectral images for all prediction samples was predicted by PLSR and displayed in different colors for visualization based on pseudo-color images with the aid of an IDL program, displaying fish freshness status and distribution vividly.(3) The Texture Profile Analysis (TPA) properties and water holding capacility for turbot flesh was determined based on visible and short-wave near-infrared spectroscopy. Three TPA properties of adhesiveness, springiness and cohesiveness was measured for turbot under chilling storage, and drip loss was used to describe the water holding capacility of frozen turbot. The best spectral pretreatment method was selected, and a novel method called Random frog was employed to select the effective wavelengths from the full-spectrum. Then PLSR, BP-ANN and LS-SVM models were established based on the effective wavelengths for determining the different physical parameters. The highest prediction Rp for adhesiveness, springiness, cohesiveness and drip loss were0.9094,0.8754,0.8462and0.8678, respectively.(4) Mapping of fat and moisture distribution in Atlantic salmon was achieved using long-wave near-infrared hyperspectral imaging. The extracted mean spectra from the hyperspectral images were correlated with their corresponding fat and moisture contents using PLSR, BP-ANN and LS-SVM. High performances were obtained by the linear PLSR models with Rp of0.9263, RMSEP of1.2405for fat prediction, and Rp of0.9366, RMSEP of1.0579for moisture prediction. Then the PLS models were applied pixel-wise to the hyperspectral images of the prediction samples to produce pseudo-color chemical images for visualizing fat and moisture distribution along the whole fillet. Besides, the feasibility of using the proposed technique to map fat and moisture distribution for whole fillets was also explored.