New Approach To Detect Freshness Of Pork Using Spectral Imaging

Although freshness is one of the top concerns to consumers, no systems are currentlyavailable to the pork industry that could quantitatively predict its spatial distribution in a rapidand nondestructive way. An attempt was made with the invention of a spectral imaginginspection system based on visible/near-infrared (VIS/NIR) acousto-optical tunable filter(AOTF) to inspect pork for the spatial distribution of freshness.Based on the analysis of the mechanism of the inspection for the freshness of pork usingspectral imagery, it was premised that the eligible muscle region of interest (EMROI) for thefreshness inspection evaluated according to the chemical index of TVB-N of pork loins(longissimus dorsi) that were air packaged and chill stored be limited to the muscular tissuewithin the5mm layer from the top surface.An AOTF-based VIS/NIR spectral imaging system was constructed using an AOTF unit,an area camera, an illuminator of quartz tungsten halogens, and a computer. Athree-tier-architecture was used for the design of its software, which was developed on theWindowsTMplatform including a control module coordinating the principal components and animage-processing module coded with HalconTM. The whole imaging system was furthercalibrated and its key specifications were determined that the optimal signal-to-noise ratio was8.02when working on the wavelength range of575~940nm and an8-min-warming was usedto stabilize the system on startup. Consequently, the VIS/NIR diffuse reflections from the topsurfaces of43subjects were spectrally imaged with the system.A method was developed to extract the EMROI from the spectral images of pork. Thecomparative analysis of the manually sampled local91-waveband-reflectance-spectra from theEMROI and from the non-EMROI in the spectral images showed that the former having anabsorbance peak at575nm. An attempt was made for the explanation of this observation. Fromthe perspective of the theory of color space, the observed absorbance peak at575nm could beexplained as the lack of yellow in the reflection, having the visual effect of adding itscomplement color, which is the blending of purple and red. This makes sense because that freshmuscle tends to turn purplish when exposed to atmosphere with little oxygen and reddish toatmosphere rich in oxygen, while purple or red is rarely shown on fat. The observed absorbancepeak could also be interpreted from the perspective of biochemistry as the combination of thoseat557nm of deoxymyoglobin and at582nm of oxymyoglobin when taking into account thatthe imagery system had a half-peak-bandwidth of20nm. The reflectance images at575nmwere used accordingly for the auto-segmentation of the EMROIs.The new method of generating signature spectra of subjects from the EMROIs wasevaluated through the modeling of the chemical measurements of TVB-N using partial leastsquare regression (PLSR). A dimension reduction algorithm was developed for this purposethat took the advantage of the relatively wide bandwidth of the AOTF-based imagery bycalculating the contribution of each waveband. The application of the algorithm improved the precision of the models while reducing their computational complexity. Various techniques ofspectral processing were evaluated in order to improve the precisions of the models, whichwere measured by the root-mean-square-errors (RMSECV) and coefficients of determination(R2CV) in leave-one-out cross validations. The result of the evaluation of standard normalvariate (SNV) for a global spectral processing on the signature spectra in reflectance showedthat it worked the best with the PLSR modeling with the9wavebands whose centerwavelengths were575,600,615,705,765,825,885,915, and935nm, respectively, withRMSECV=1.94mg/100g, R2CV=0.89but the spectral images for a prediction required anacquisition lasted for70s. In order to speed up the prediction, Savitzky-Golay (S-G)-basedderivative and smoothing filters were also evaluated for a local spectral processing on the samespectra, and the results showed that the7-point S-G second derivative filtering served the bestfor a local spectral processing when working with the PLSR model with4wavebands centeredrespectively at610,625,650, and900nm. The precision achieved by the local model could notcompare to that by the global model though, sufficient for the purpose of freshness inspectionwith RMSECV=2.71mg/100g and R2CV=0.79. But due to the reduction of the number ofwavebands, the acquisition of the spectral images needed for a prediction was shortened to17s.A pixel-wise prediction was introduced to evaluate the spatial distribution of the freshnessover sample surfaces. For this purpose, a method for the validation of pixel-wise predictionswas created and used on the practice based on the global spectra prediction model. Resultsshowed that the mean prediction value over the sample surface coincided well (RMSECV＝2.58mg/100g, R2CV＝0.81) with the corresponding chemical references. The introduction of thevalidation of spatial predictions changed the practice of the art of applying models developedfrom the spectra of regions directly to the spectra of pixels, indicating that it was desired that aproper spatial averaging be applied for de-noising prior to a pixel-wise prediction. Acomparative analysis was carried out to demonstrate that the new approach would help deliverpixel-wise predictions with sound statistical meaning.The freshness distributions over sample surfaces were presented with pseudo-color imagesconfigured for the best usability to transit from blue, yellow, orange, to red, showing intuitivelythe distribution of regions on different deterioration stages from very fresh, fresh, less fresh tospoilt. The histogram analysis of all the pixel-wise predictions showed that the mean of thepredictions over a sample surface coincided well with its chemical measurement, indicating thatthe introduction of EMROI helped to keep the regions affected by the specular reflectionoccurred on the subject surface during acquisition out of the modeling procedure, safeguardingits precision.The practice of the spatial analysis of the spectral information of this research showed thatthe technique of quantitative visualization is a new approach to the examination of the qualityattributes of meats by exploring the chemical changes of different components and theinteraction of their products during storage. Industrial relevance could also be seen in the newmethod and system that the spectral imaging system developed in this work is capable topredict rapidly and precisely the spoilage indicating chemicals of TVB-N with the unprecedented presentation of freshness distribution over sample surfaces in pseudo-color,visualizing vividly the inherent spatial heterogeneity of the deterioration over pork surface.Acquisition of the spectral images is carried out statically, with no dependency on a conveyorsystem，offering a promising prospect of the mobile inspection in situ at reception sampling oreven at marketplace.