Study On Biospeckle Activity And Detection Of Potato Invisible Damages

Potato is one of the four kinds of staple food around the world.With the advancement of the"Potato Staple Food Strategy" in China,mechanization will stand out in potato production,which will bring about the occurrences of mechanical damages during harvest and postharvest handling of potatoes.Therefore,how to monitor mechanical damages is a vital project.In this study,finite element analysis(FEA)was firstly combined with image processing technique to create potato model under drop case.Occurrence patterns of damages were studied using FEA and drop test.Specific relationships between damages and biospeckle activitity(BA)were investigated,and modified processing methods were proposed.Finally,damages were recognized using machine learning methods,along with BA,color and texture features.Research contents and main conclusions were listed as follows.(1)Establishment of FEA model for potato under drop case based on image processing technique Two image processing-based algorithms were proposed,in order to aid in building potato geometric model and obtaining material properties,which will ensure the accuracy of FEA model.Feature points on contour lines were obtained to constitute feature lines 1,which can determine crosswise structure;feature lines 2 were drawn based on feature lines 1,which can determine lengthwise structure.An algorithm aiming to calculate radius of curvature on contact points was proposed,which enables the calculation of(time-depended)modulus of elasticity.Consequently,elastic and viscoelastic properties were determined.Two image processing-based algorithms were both proved to be reliable.(2)Dynamic response and damage behavior analysis for potato collision based on FEA FEA model,along with drop tests were used to investigate dynamic responses during collision,in order to unveil damage occurrence patterns.Potato model in FEA and samples in drop test were dropped from 200,400 and 600 mm.Five response parameters of contact time Tc,peak force time Tp,maximum impact force Fmax,maximum deformation Dmax and restitution coefficient r were extracted and compared.Results from FEA showed that,Tc and Tp decreases slightly,Fmax and Dmax increases apparently with the increase of drop height,while r remains high regardless of the drop height.Compared with elastic model,results derived from viscoelasitc model were more realistic.Damge bahavior were monitored treating potato as viscoelasitc model.Ocurrence of damages were influenced by fators in priority order:material property>impact angle>drop height>mass.Tc was found to be negative linear correlated with maximum stress,which can be used to monitor the happening of damages.Results also showed that skinning injury can be easily happened,and internal bruise occurs in subsurface area at first.(3)Configuration and optimization of biospeckle imaging system Biospeckle imaging system was configured.Hyperspectral imaging technique was utilized to determine the optimal color to detect invisible damages.Acquisition parameters were optimized according to contrast of one speckle pattern,which can be assessed from multi-scale leves.Firstly,system elements were chosen according to research object/index;secondly,imaging and capturing parameters were optimized;finally,speckle images were assessed to validate the system effectiveness.Hyperspectral imaging technique was utilized to determine the optimal color to detect skinning and subsurface damages.Red color was shown to perform better,which can provide theoretical support for selection of laser light.Parameters such as object distance,size of ROI,exposure time were adjusted to meet the requirements of speckle size,saturability and light uniformity.(4)Modification of processing methods for speckle time-series signal Modified methods were proposed according to the shortcomes of exsiting time-domain IM and frequency-domain wavelet entropy(WE)method.Outliers detection was added in IM,and correlation of contiguous speckle patterns(CCSP)was proposed.Scale enhancement factor was introduced into WE,creating modified wavelet entropy(MWE).Six different processing methods were involved in this study,including THSP_IM,CCSP_IM,THSP_WE,THSP_MWE,CCSP_WE and CCSP_MWE.Optimal methods were chosen according to application scenarios based on resolution,stability and robustness.Modified methods performed better compared with former ones in this study.(5)Specific relationship between potato damages and BA Based on FEA damage model of potatoes,BA,color and texture features were extracted and compared,to explore the variation patterns between damages and speckle patterns.For skinning injury,RGB images and speckle patterns were collected to extract overall 16 features,including color,texture and BA features.BA features were obtained by optimal methods of CCSP IM and THSP_MWE,and they were used to monitor activity changes.Results showed that activity was only related with storage time among storage time,sample mass and injured parts.BA increased significantly within 1 d,then decreased till approaching normal skin.For subsurface damage,color,texture and BA features were extracted.The preferred methods CCSP_WE and CCSP_MWE were used to monitor the activity profile of normal and damaged samples.No significant differences were found in BA of normal,slight bruise and blackspot bruise,and BA did not depend on the depth of damage.The effect of texture features on BA was significant,and the more obvious the texture,the lower the BA.(6)Qualitative/quantitative analysis of occurrence of subsurface bruise Potato may present normal,slight bruise and blackspot bruise after impact.The distribution of theses damages were analyzed qualitatively,depth d1 and height d2 of blackspot bruise were also analyzed quantitatively.Influence of mass m,radius of curvature r,curvature ratio Ratio and texture feature Con on liability of damages,d1 and d2 were investigated.Results showed that m had no influence on occurrence of damages,liability of occurrence for different parts in order was:stem-end>bud-end ? body.Damages on stem-end were prone to blackspot bruise.r and Con had significant influence on damage occurrences(P<0.05).The smoother the surface,and the smaller the radius of curvature,the more prone to damage.The four factors had no significant effect on d1(P<0.05),but d1 was significantly smaller at the stem-end than the other two parts.The effects of m and r were statistically significant on d2(P<0.05).The larger the mass and the smaller the radius of curvature,the greater the height of the damage.(7)Skinning injury detection based on BA In view of the variation patterns between skinning injury and BA,classifiers were constructed.Color and texture features were combined afterwards,in order to recognize skinning injury.Classifiers of logistic regression(LR),decision tree(DT)and support vector machine(SVM)were used in dichotomous classifications.Results showed that LR performed better than DT and SVM.BA contributed better classification results within 1 d,while color features were better after 1 d.BA and color features were combined to create the classifier of Fusion_LR.Classification accuracies for normal skin and skinning injury were 89.76%and 92.76.Overall classification accuracy was 91.28%.(8)Subsurface bruise detection based on BA In view of the variation patterns between subsurface damage and BA,classifiers were constructed.Color and texture features were combined afterwards,in order to recognize normal,slight bruise and blackspot bruise.Classifiers of LR,DT and SVM were selected to classify the subsurface damages of potatoes on different parts.LR and DT were used for the dichotomous(normal-damage)and SVM for both the dichotomous and the tripartite classifications.Results showed that optimal classifiers for bud-end,stem-end and body were BA_DT,BA_LR and Fusion_DT,respectively.The classification accuracies of normal samples were 47.8%,41.7%and 53.3%;the classification accuracies of damaged samples were 70.8%,84%and 73.3%;and the overall classification accuracies were 61.2%,78%and 66.1%,respectively.Fusion_LR was the preferred classifier when ignoring specific potato parts,87.1%of damaged samples were correctively recognized,but only 34%of normal samples were correctly identified with an overall classification accuracy of 70.9%.In the tripartite classification case,the preferred classifiers for different parts were Fusion_SVM,but none of the classifiers had the ability to recognize slight bruise.