Study On The Purity Identification Of Maize Seed Based On Near Infrared Imaging

Maize seed purity is one of the key factors to limit its production. It can promote grain production and farmers’ income to improve the quality of seeds by purity identification, and it’s of great significance to ensure national food security to improve the purity of maize seed. Method of identifying maize seed purity based on near infrared imaging was studied in this paper. Near infrared spectroscopy qualitative analysis software of seeds was developed. The research content is summarized as follows:1. Purity of six maize seed combinations (each combination contains a hybrid and its female parent) were identified using spectral information and image information in hyperspectral image. The average accuracy of seed purity identification model based on spectral information has met the performance requirements of seed sorting (three combinations achieved accuracy above90%, the other three combinations yield accuracy more than80%). All of the accuracies of six combinations’ purity identification models based on shape and texture characteristics of single band image were above80%, which can satisfy the actual request. Four wavelength ranges (900-1000nm,1100-1200nm,1250-1350nm,1400-1500nm) which can reflect the difference between characteristics of hybrid seed and its female parent correctly and stably were determined.2. Method of gathering images of four characteristic wavelengths (910nm,1100nm,1350nm,1450nm) based on near infrared camera and band-pass filters was studied. Two kinds of image acquisition mode (reflection and transmission mode) were explored and compared, and transmission mode, through which hybrid seed and its female parent can be distinguished easily, was found to be better, and the difference between images of one seed’s embryo side and nonembryo side can be overcome by transmission mode. The performance of purity identification model based on transmission mode of five maize seed combinations were tested, and two combinations yielded accuracy above90%, the other three combinations achieved accuracy above80%. Stability of Nonghua101purity identification model was tested by seven sets of test data gathered in different environments, and all of the accuracies were around95%. The performances of models have satisfied the practical needs.3. Near infrared spectroscopy qualitative analysis software of seeds was designed and implemented to improve the efficiency of the qualitative analysis, and the reliability of results. Based on the evaluation indexes such as separability, and departure degree, the software can forecast the performance of qualitative analysis models of a batch of data to be classified. It can also guide the selection of data preprocessing and feature extraction methods, and determine the optimum data dimension of feature extraction automatically. The software takes both the model’s classification ability and generalization ability into consideration, and enhances the accuracy and robustness of the qualitative analysis model by improving data standardization and Biomimetic Pattern Recognition algorithms.