| As an important production resource,the quality of maize seeds determines the success or failure of planting,and it also has an important influence in the increase of farmers’ income and the steady development of rural economy.It is not only a key point of ensuring the quality of seeds,but also coudle improve the yield of products,ensure the safe storage and transportation of seeds,accelerate the process of seed industrialization.In this paper,the internal and external quality of maize seeds were studied,and we aimed to develop algorithms to inspect the quality of maize seeds by using computer vision combing with hyperspectral imaging,and chemometrics methods.The proposed systems and methods could provide a potential support for the quality detection of maize kernel with multi-spectral technique.The main research conters and results are listed as follows:(1)Detection of MC in maize seed with 400-1000 nm hyperspectral image was studied.Firstly,hyperspectral images including both front and reverse side of maize kernel were acquired,spectral data in centroid region was extracted,and competitive adaptive reweighted sampling algorithm(CARS)was used for characteristic wavelength selection.And then prediction models including front side,reverse side and mixture location prediction model were built for MC prediction.Secondly,spectral curve in different parts of hyperspectral images were contrasted mutually to judge if the maize kernel appeared in the image was front side upward(embryo up)or not,and two wavebands(520 nm,560 nm)were selected for front and reverse side detection with band math.Finally,MC of 45 validation set samples were detected with the algorithm proposed in this paper.Results show that accuracy of front side and reverse side detection is about 97.8%,100%.The detection results of MC prediction models with spectral information at different positions(front side,reverse side,mixture location), are 0.896,0.948,0.885;RMSEV are 0.823%,0.593%,0.858%.PLS model established with spectral information of single side is better than the model established with spectral information of mixture location,and the model established with spectral information of reverse side(embryo)is better than the model established with front side(endosperm).(2)Detection of MC in maize seed with 1000-2500 nm hyperspectral image were studied.Hyperspectral images including both front and reverse side of maize kernel were acquired for analysis.Prediction models including front side,reverse side and mixture location prediction model were built for MC prediction,and four wavebands(1104 nm,1304 nm,1454 nm,1751 nm)were selected for front and reverse side detection with band math.Results show that accuracy of front side and reverse side detection is about 97.8%,100%.The detection results of MC prediction models with spectral information at different positions(front side,reverse side,mixture location), are 0.969,0.946,0.947;RMSEV are 0.464%,0.616%,0.667%.PLS model established with spectral information of single side is better than the model established with spectral information of mixture location,and the model established with spectral information of side(endosperm)is better than the model established with reverse side(embryo).(3)The application of machine vision in the detection and sorting of maize seeds was studied.Effective feature parameters selected by I-Relief algorithm was used to train classifiers for seed identification and detection.Result shows that if whole surface color feature parameters were used to train classifier for the detection of moldy maize seed,the accuracy could reach 98% using naive bayes classifier.If color feature parameters of seed tip were used to train classifier for the detection of moldy maize seed,the accuracy could reach 99.3% using naive bayes classifier.If shape feature parameters of maize seed were used to train classifier for the detection of damaged maize seed,the accuracy could reach 97.3% using BP neural network. |
PDF Full Text Request