Real-time Detection Of Kernel-like Impurity And Unsound Kernel In Wheat Using Machine Vision

China is the largest producer and consumer of wheat in the world, and the inspection of wheat quality is very important for the grain industry. Up to now, the inspection of Kernel-like Impurity and Unsound Kernel in wheat is rely upon manual method still, which is neither objective nor efficient. Machine vision provides a rapid and objective means for evaluating the appearance quality of wheat. The research integrates knowledge of optics, hyperspectral, image processing and pattern recognition to analyze the real-time detection of the Kernel-like Impurity and Unsound Kernels in wheat, and a prototype was explored utilizing with technologies of mechanism, electronic and computer. The main contents are as follows:1. Hyperspectral experiments and image acquisition system. Hyperspectral images were used to determine the optimal band according to each band's contribution to the most effective principal component. The study has found that the contribution of 630 nm band is the most useful band for morphological feature, which is right close to the wavelength of the red light used in this thesis.2. Real-time wheat image preprocessing and feature extraction. Through the image histograms and the contrast analysis of wheat under different color backgrounds, the black and opaque background turns out to be the best one for wheat identification. After pretreatment and some image process steps, including unnecessary background cutting, homomorphic filter, image threshold segmentation, morphological image processing, color space conversion and wavelet transform,7 morphological features,27 color features and 144 texture features were extracted.3. Comparison of different pattern recognition methods of Kernel-like impurity and Unsound Kernel in wheat. The samples include 1169 Sound Kernels,897 Kernel-like Impurities,710 Black Germ kernels, and 627 Damaged Kernels. There are several analysis algorithms used for data processing, such as genetic algorithm and the support vector machine (GA-SVM), principal component analysis and support vector machine (PCA-SVM), partial least squares discriminant analysis (PLS-DA), partial least squares and support vector machine(PLS-SVM), principal component analysis and artificial neural network (PCA-ANN), linear discriminant analysis(LDA). According to the research, combining genetic algorithm and the support vector machine (GA-SVM) is the best method, the accuracy of which was 99.34%; PCA-SVM gets the best accuracy:97.2%,98.4% and 97.9%for Black Germ Kernels, Broken Kernels and Sound Kernels. 4. Prototype development for real-time detection of Kernel-like impurity and Unsound Kernel in wheat. The mechanical part mainly includes the main stents, the single graining device, the transmission device and the material feeding device. The circuits section of the machine (lower computer) consists of core chip of ATmega128, communication circuit chip and some peripheral circuits. According to the image processing and pattern recognition method determined in experiment, the code is realized with VC++ software. The Personal computer (upper computer) can control the camera setting, image acquisition, image processing, feature extraction, pattern recognication and communication directly. The hardware and software are collaboratively debugged, and finally all software used is packaged into a setup.exe bag. Users double-click setup.exe to complete software installation and then it will be ready to work.