| Fusarium head blight(FHB),a fungal disease caused by Fusarium graminearum,is one of the main diseases in wheat growth.FHB mainly breaks out in warm,humid a nd semi-humid areas,and the temperature and climate change will accelerate the occu rrence of scab around the flowering stage.The disease can cause a serious decline in wheat production.Moreover,excessive intake of foods infected with FHB can cause s ymptoms such as dizziness and vomiting.In severe cases,it can cause disorders of the gastrointestinal and nervous system,and even become cancerous.Therefore,it’s urge nt to prevent and control the wheat FHB.At this stage,there are basically two method s to detect the wheat FHB: manual,or chemical detection.Manual detection is time-c onsuming,laborious and subjective,and long-term human eye observation can cause ocular muscle strain and confuse healthy and diseased grains so that the manual detect ion efficiency and judgment accuracy would be reduced.Chemical detection is mainly consisted of gas chromatography(GC),high performance liquid chromatography(HP LC)and enzyme-linked immunoassay(ELISA),etc.Though the chemical approach is relatively accurate,there are many problems such as damage to samples,complicated operation steps,excessive detection time and environmental pollution.Therefore,in o rder to ensure food safety,it’s necessary to develop a method for rapid,accurate and n on-destructive detection of wheat FHB.In this paper,the wheat kernels infected with FHB and healthy wheat kernels wer e taken as the research samples.Scanning electron microscopy(SEM)was used to obs erve the microstructure changes of healthy and diseased samples.Physical and chemic al analysis experiments were used to collect the indicators of moisture,protein content and hardness of healthy and diseased wheat kernels.The healthy and diseased sample spectrum data from 400-4000cm-1 were collected to analyze the influence of microst ructure and internal content changes on the spectrum.The wheat kernel hyperspectral image data(400-1000nm)were acquired to select the feature wavebands.Different m achine learning algorithms based on feature wavebands were used to establish the rec ognition models of FHB wheat kernels,then the optimized model was selected by the comparison of model accuracy.Different deep learning algorithms based on feature w aveband fusion images were used to establish the recognition models of FHB wheat k ernels,and the optimized model was chosen according to the model accuracy.The em bedded deep learning real-time detection system was built to realize the online real-ti me detection of FHB wheat kernels.The main research contents and conclusions of this paper are as follows:(1)Study on the influence of changes in the microstructure and internal com ponents of healthy and diseased wheat kernels on the spectrum.The wheat flour samples were observed by SEM under 500 and 1000 magnifications.Inside the healthy samples,the starch granules were tightly bound,and the combination of starch granules becomes looser as the content of deoxynivalenol(DON)increased,could be observed.The wheat kernels were observed by SEM under 30,40,200 and 500 magnifications.In the healthy kernels,the surface was rounded and compact,could be observed.In the FHB kernels,the surface was gray and shriveled,which appeared pits and holes,and the texture was soft and easily damaged.The aleurone layer inside the diseased kernels was destroyed and many cell walls disappeared,that appeared to a large number cavities full of hyphae.Damage to the surface of macromolecules such as protein and starch resulted in a honeycomb structure,and there were many finely divided starch grains inside,could be observed.With the deepening of infection,the surface of wheat grains was covered with mycelia,and the colony matrix showed red,and part of the colony matrix showed dark red,leading to the appearcnace of red head on the surface of wheat grains.The national standards of GB/T 21305-2007,SN/T2115-2008 and GB/T 21304-2007 were referred to detect the moisture,protein content and hardness of wheat kernels.According to the detection results,there were significant differences in water content,protein content and hardness between infected and healthy wheat grains(P<0.05).By comparing the Fourier transform infrared spectroscopy data of healthy and infected wheat flour samples ranged from 400-4000 cm-1,the absorbance of protein feature band ranged from 2168-2180 cm-1 had a falling-off,and the absorbance of protein characteristic band related to N-H stretching vibration in the range of773-1020 cm-1,1500-1530 cm-1 and 2050-2060 cm-1 all decreased.Therefore,after wheat grains were infected with scab,the content of crude starch and crude protein decreased,and the change of protein content would cause the change of Fourier spectrum.The changes of microstructure and internal composition of wheat grains infected with scab will affect the information response of hyperspectral images,which provides theoretical support for subsequent identification of scab grains based on hyperspectral image technology.(2)Study on the recognition methods of FHB wheat kernels based on spectral information and machine learning.The region of interest from wheat kernel hyperspectral images was selected for further spectral processing by ENVI 5.3software.Multiplicative scatter correction(MSC)and Standard normal variate transformation(SNV)algorithms were used to preprocess the reflectance of spectrum.Successive Projections Algorithm(SPA)and X-Loading Weights(X-LW)were used to select the feature wavebands of the processed spectrum data.Support vector machine(SVM)and Partial least squares discriminant analysis(PLS-DA)were used to create the models.The accuracy of the training sets of the sample population was above 80%,and the accuracy of the test set was also above 80%.Under the same preprocessing method and feature waveband selection method,the recognition accuracy of the model established by SVM algorithm was generally higher than model established by PLS-DA algorithm.Moreover,SNV+SPA+SVM had established the best model with the feature ban ds of 733,540,and 668 nm,and the training accuracy reached 97.31% and the test acc uracy reached 96.42%.(3)Study on the recognition methods of FHB wheat kernels based on hyperspectral feature wavelength fusion images and deep learning.Three feature wavelengths of 773 nm,540 nm and 668 nm were selected and combined into feature wavelength fusion datasets.The size fusion images were adjusted by means of sequential cropping,and fusion image datasets were made into 416×416 and1024×1024 pixel image datasets,respectively.The 416×416 and 1024×1024 pixel image datasets were used to train the YOLO v3 and U-Net models,respectively.The parameters of learning rate and batch size were used to optimize the YOLO v3 and U-Net models.The results showed that U-Net network had the best effect when learning rate was set to 0.001 and batch size was set to 8.When the iteration was set to 3000,the optimized U-Net model reached highest accuracy of 96.1%.(4)Study on the real-time detection system based on embedded system and deep learning method.In order to realize the online real-time detection of FHB wheat kernels,an assembly line was designed,including separation modules,conveyor belt,acquisition modules and control modules to complete the grain-by-grain transmission and rapid identification of wheat kernels and output the detection results.Two diving plates in separation modules,which groove widths were set to 6 mm and 5 mm,were used to separate the piled and sticky wheat kernels.The RGB image datasets,which collected by the industrial camera,were used to train the U-Net model.When learning rate was set to 0.001 and iteration was set to 3000 and batch size was set to 2,the U-Net model was optimized and the model had the highest accuracy of 95.78%.Combined with the optimized U-Net model and hardware system,an online real-time detection of FHB wheat kernels could be fulfilled. |
PDF Full Text Request