Diferansiyel Hiperspektral Görüntüleme Tekniği ile Kimyasal Madde Tespit Yöntemlerinin İncelenmesi

Abstract

The aim of this thesis is to investigate the performance of some classification algorithms for differential spectral reflections of different chemical materials on different background materials. In the study, firstly, differential spectral reflections have been calculated for the reflections, that have been collected in different wavelengths by the active camera system, of different substances and materials that move with conveyor belt. On these noisy reflections, different parameters of Savitzky-Golay Filter, Moving Average Filter, Gaussian Filter and Median Filter algorithms have been studied and the results were examined. As a result of filtering, the classification algorithm is used to classify the filtered data with the selected filter type and selected parameters. For the Support Vector Machines algorithm in the feature-based classification category, a feature extraction method, Principal Component Analysis, is applied. Then, Support Vector Machines are examined with different parameters for the selected features by Principle Component Analysis method and the results of Support Vector Machines with the best parameters according to performance are analyzed. In the signature-based classification category, Matched Filter and Spectral Angle Mapper algorithms are used. The performances of all classification algorithms were compared and analyzed the good and bad sides of each other. Then, in order to increase the performance of classification algorithms by decreasing the number of false positives, the final processing algorithms developed in this thesis have been used. First of all, it has been observed that each substances according to their chemical characteristics regardless of the background material had an increase or decrease in the differential reflection in a certain wavelength range, and the Slope Thresholding algorithm which uses this information has been developed. After the classification process, the Slope Thresholding algorithm which provides the slope of this increase or decrease in the selected wavelength range to be higher than a certain threshold value, has been applied on data, and it is observed that the number of false positives decreased without a significant decrease in the number of true positive number as expected. The Pixel Removal algorithm, which is another final processing algorithm and which allows the removal of a few errors on the background data, has also been developed within the scope of this thesis. As a result, in this thesis, it has been demonstrated that chemical substances could be detected by using differential hyperspectral imaging technique and different algorithms that provide this detection have been compared.