Research On Detection Method Of Maize Variety Based On Hyperspectral Image

Maize is one of the world’s three major food crops.Ithas a strong adaptability,wide planting area,large yield,wide range of uses.It has extremely important strategic significance for China’s national economic development and social stability.At present,the maize seeds are highly commercialized,and the variety is very confusing,which brings great inconvenience for the farmers to ease the purchase and market supervision.Study on variety authenticity detection technology which can meet the requirements of market supervision,is rapid and nondestructive is conducive to the protection of maize seed market and norms.In this study,rapid and nondestructive varieties of maize seed authenticity detection technology as the research object.This study discussed the preprocessing problem,the feature selection problem,the recognition and detection problem of hyperspectral images applied to the authenticity detection of maize varietiesusing hyperspectral image processing and pattern recognition technology.This paper puts forward suitable for practical application of denoising,segmentation,and the methods of feature selection and recognition detection.The main research contents and conclusions are as follows:(1)In this paper,a hyperspectral image denoising model based on Contourlet transform and threshold function is proposed,and a general image segmentation method based on histogram slope difference adaptive threshold is proposed.For the problem of the noise of the maizeseed in the practical application of hyperspectral image,this paper combined with the practical application to construct a denoising model based on threshold function on the basis of analyzing the principle of Contourlet transform.At the same time,this paper through the introduction of edge preserving coefficient analysis experiment shows that the model has better denoising effects on hyperspectral images of maize seeds with different intensity(variance)noise,and outperforms the mean and wiener filtering methods.According to the practical application of maize seed hyperspectral images due to noise and not easy segmentation,especially automatic technology,the bottleneck problem of robust image segmentation based on histogram,distribution characteristics of slope experiment found the difference,proposed a new image segmentation method based on gray scale transform from the distribution of labeled pixel space conversion to grayscale distribution density in order to realize the method(DGTS).According to the analysis of the syntheticimage and real maize hyperspectral image experiments,this method compared to the EM and K-means methods are more accurate and effective to distinguish between different gray value and characteristics of image segmentation.Moreover,it has better adaptability and robustness in segmenting and processing hyperspectral images containing noisy maize seeds,and is more suitable for detecting the actual application.(2)This paper presents characteristics of maize varieties bandratio maximum load method and varieties selection based on morphological characteristics,selection of texture feature extraction methods,and puts forward the feature fusion parameters to achieve a variety of detection method,to solve the image of maize hyperspectral dimensionality reduction and "Homogeneous spectrum" problem.Based on the analysis of PCA transform method,the characteristic band selection method based on the corresponding maximum load factor of each principal component after PCA transformation is proposed.According to this method,the three bands of blue,green and near infrared are selected 5 bands were used as the characteristic bands of maize seed varieties and analyzed their spectral characteristics.Based on the comprehensive consideration of maize seed morphological characteristics,the morphological parameters of five maize varieties with basic geometrical and invariant moments were selected.The extraction methods of these five morphological parameters were discussed and analyzed.The algorithm for calculating the grain circumference by using 8-chained code is also described in detail.Based on the comprehensive analysis of maize seed texture characteristics,the four gray-level symbiotic matrices eigenvalues of maize varieties were selected as texture features,and their extraction methods were expounded and the characteristics of texture features were analyzed experimentally.The above research has laid a foundation for the optimization and in-depth study of hyperspectral image recognition based on multi-character maize seed.(3)Based on the identification and detection of maize seed hyperspectral images based on multi-class features,this thesis demonstrated the morphological characteristics of maize seed and the classification and recognition ability of maize seed in different bands,and proposed the fusion of many characteristics of maize seed hyperspectral Method of image classification recognition.Based on the PLSDA method,the classification and recognition of maize seed hyperspectral images based on morphology,texture and multi-class characteristic parameters of 10 kinds of maize varieties in multi-band,full-band and characteristic bands were studied and analyzed.The average recognition rate is 64.68%,the highest average accuracy is 49.64,and the average recognition rate is lower.The average training and testing of the whole band is in the multi-band and full-band range.(97.77%and 90.80%)were significantly better than the average training and testing accuracy(83.89%and 73.78)of the multi-band.Under the whole band,the different texture parameters had different ability to discriminate the maize seed variety,and the energy,mean value and entropy The combination of three texture features achieves the training and testing of the highest classification accuracy of 98.87%and 98.21%.In the characteristic band,energy,mean,entropy 3 texture feature combination training and testing accuracy can only reach 91.05%and 84.76%,but the fusion characteristic parameters,the ideal state of training and testing accuracy can reach 99.37%and 98.14%.It shows that the reduction of the number of bands will reduce the accuracy of the classification and recognition of the model,but the fusion shape features and texture features can make up for the band to a certain extent,reduce the loss caused by the information.It also proves the correctness and validity of the proposed fusion of multi class features for maize seed variety detection.(4)This paper discusses the realization of hardware and software design of maize breed detection prototype system using hyperspectral image.Based on the demand of maize seed market management,the hardware and software modules of prototype system are planned and designed by object-oriented method.The hyperspectral acquisition equipment was selected and the integrated box and control system modules were designed.The software functions of the prototype system are planned and designed.The timing and relationship between the use cases and the use cases of the software function modules of the prototype system are analyzed and some modeling is carried out using the modeling tools.Finally,from the perspective of the management of prototype system construction and application,the establishment criteria and norms of maize seed hyperspectral standard feature database are discussed.