| Abstract [eng] |
The essence of this project is to optimize a combination of methods capable of effectively recognizing and detecting microplastic particles in unevenly illuminated microscopic images. In this work, binarization methods were applied with the aim of differentiating microplastic objects from the background of the images. After the binarization process, it was observed that both the Otsu and Sauvola methods fragment microplastic particles into separate parts due to the uneven brightness and contrast in the microscopic images. Additionally, the gray spots from the filter cause an excessively high number of falsely detected microplastic particles. However, the Sauvola method is more effective because it is able to identify a larger number of true microplastic objects, despite also detecting a considerable number of falsely identified particles. After the binarization and segmentation of particles, morphological decisions were applied, which significantly improved the detection results by reducing the number of falsely detected particles. Further refining the results obtained from the methods, a machine learning classification approach was implemented. Initially, a solid foundation for the machine learning process was established by performing an analysis and selection of features from the microplastic dataset, tailored for machine learning classifier models. Considering the evaluation of classifier models and binarization based on various metric indicators (precision, recall, and F1 score), the machine learning solution helped to more effectively recognize microplastic particles. The most optimal classifier model tested in this project is the random forest classifier, applying the hard negative mining method. This additional approach enhanced the classifier’s precision in identifying true microplastic particles. |
