The Study On Recognition Methods Of The Contamination On Poultry Carcass Based On Hyperspectral Imaging

In the process of mechanized poultry slaughtering,poultry carcasses are highly susceptible to the contamination of intestinal feces,bile and blood.Some residual contaminants will be diluted by water and become poorly visible.It is difficult to identify these contaminants by artificial detection or conventional machine vision.The contaminated carcasses are hazards to food safety.Therefore,more effective detection methods must be developed to detect contaminated poultry carcasses.The contaminated carcasses should be removed or be carried out appropriate treatment according to the results of detection to ensure the safety of poultry meat products.Due to hyperspectral image possesses both spatial information and high spectral resolution,it has been widely used in the research field of detection and become a powerful tool for intelligent production.Based on the technology of collecting,analyzing,modeling and image processing of hyperspectral image data,a set of detection methods for feature recognition of contaminants on chicken carcasses was proposed in this study.Three detection algorithms were proposed:? the contaminated regions were modeled identified by successive projections algorithm(SPA)-multivariate linear regression(MLR)-receiver operating characteristic analysis(ROC)classifier;?the identification method of diluted contaminants based on the two-dimensional spectrum of hyperspectral data and the two-dimensional scatterplot;? the segmentation method of the adaptive region of interest(ROI)was constructed based on the regional features.The charactristic band with the maximum recognition degree of contaminated regions was used to realize the intelligent recognition of contaminants with low concentration and low visiblity on chicken carcasses.The main contents and results are as follows:1.The SPA-MLR-ROC classifier was developed,which can extract features of the detection target modelally.It can obtain high true positeve rate(TPR)and low false positive rate(FPR)based on abundant spectral information.TPR is the ratio of the number of actual contaminated regions,wich are also correctly determined by the classifier,to the number of all contaminated regions.FPR is the ratio of the number of the non-contaminated pixes,which are identified by the algorithm as contaminated pixes,to the number of all non-contaminated pixes.The algorithm flow is:the hyperspectral images which were used for training and validation were obtained by a hyperspectral imaging system.The regression model of the classifier was established by MLR based on twelve characteristic wavelengths(505,537,561,562,564,575,604,627,656,665,670 and 689 nm)selected by SPA,and the optimal threshold T=1 was obtained by the ROC analysis.The SPA-MLR-ROC classifier provided the best detection results compared with SPA-partial least squares regression(PLS)classifier and SPA-least squares supported vector machine,(LS-SVM)classifier.It is characterized as a precise classifier that the TPR was 100%and the FPR was 0.392%.2.The hyperspectral two-dimensional(2D)correlation spectrum was constructed and 2D scatterplots were used to detect the contaminats on the surface of chicken carcasses,which are diluted with water at a ratio 1:1.The features of the deluted contaminants in hyperspectral image become weake.Thus,it is necessary to select characteristic bands by extending the spectral dimension to 2D.The 2D scatterplot can be constructed by the characteristic bands that best reflect the features of the targets.The pixels in the contaminated regions were identified by the threshold segmentation in the 2D scatterplots.The algorithm flow is:from the collected hyperspectral data,a set of uncontaminated regions of interest(ROIs)and four sets of contaminated ROIs were selected to constitute a training set.The average spectrum of the uncontaminated ROI was treated as the original spectrum and that of each contaminated ROI was regarded as influenced spectrum.The difference between the original spectrum and the influenced spectrum is the difference spectrum.All difference spectra were used to conduct correlation analysis,from which the 2D hyperspectral correlation spectrum was constructed.Two maximum auto-peaks and apair of cross peaks appeared at 656 nm and 474 nm.Therefore,656 nm and 474 nm were selected as the characteristic bands because they were most sensitive to the spectral change induced by the contaminants.The 2D scatter plots of the contaminants,clean skin,and background in the 474-and 656-nm space were used to distinguish the contaminants from the clean skin and background.The threshold values of the 474-and 656-nm bands were determined by ROC analysis.According to the ROC results,a pixel whose relative reflectance at 656 nm was greater than 0.5 and relative reflectance at 474 nm was lower than 0.3 was judged as a contaminated pixel.A region with more than 50 pixels identified was marked in the detection graph.This detection method achieved a recognition rate of up to 95.03%at the region level and 31.84%at the pixel level.The FPR was only 0.82%at the pixel level.3.The dual-band algorithm based on regional features was developed to identified the contaminants with low concentration and low visiblity.If the contaminants on the surface of carcasses are diluted with water at a 1:2 ratio,the contrast between contaminants and skin is very small.At the same time,the false positives caused by abnormal skin surface,such as skin folds,the connections between different tissues and so on will have a significant effect on the detection results.Therefore,the chest and abdomen of the carcass,which are most vulnerable to contamination,should be extracted as ROI to decrease FPR.In the dual-band algorithm based on regional features,one band is used for masking and the other is used for detection.The ROI obtained by masking can be easily adjusted to the maximum size according to the shape and position of the carcass.The detection band was selected based on the principle that the contaminated regions had the highest recognition degree.The experimental results show that this algorithm can effectively extract the features of the samples and realize the recognition task in hyperspectral image.The algorithm flow is:firstly,the 675 nm band,in which the identity of the ROI is the best and the spectrum difference between the ROI and the edge of the ROI is the biggest,was chosen.The 675 nm band image was binarized and the mask was extracted using region growing on the biggest connected area.Then the "and" operation between the mask and the 400 nm band image with the most discriminability of contaminants was carried out.The max ROI which can self adapt according to the position and shape of the chicken carcass was obtained.Finally,the labeling method was used to recognized if there are contamination within the segemented ROI.The results showed that the max ROIs which could self adapt to the position and shape of the chicken carcass were extracted.The average correct identification rate of contaminations such as blood,bile and feces was 81.6%.4.The applicability under the irradiation of five kinds of standard light sources(halogen,D65,TL84,CWF,and U30)of the characteristic bands extraction method based on the discriminability of contaminated region was disscussed.At the same time,the results of dual band detection algorithm in two hyperspectral image systems were compared.The results show that the characteristic bands with most discriminability of each contaminated region remain consistent under the irradiation of five standard light sources,respectively.The accuracy of the dual-band detection method is also consistent in the hyperspectral system using EMCCD cameras and halogen light sources as well as in the hyperspectral system using CCD hyperspectral cameras and LED light sources.Thus it can be seen that an appropriate feature extraction method is not limited by the light conditions,and thus it better adapts to complex testing environment.The detection objects in each algorithm are not limited to a certain type of contaminant,and each algorithm can simultaneously detect all types of contaminants on the surface of the carcass.In the process of production,detection algorithm can be selected according to detection requirements.For the obvious contaminated regions on the carcass surface,the SPA-MLR-ROC classifier can be choosen to achieve high-precision detection;it is appropriate to detect the diluted contaminants by the 2D scatter plots;if the contaminants have low concentration and poor visibility,the dual-band algorithm should be used.