Özet:

One of the primary challenges in designing defect detection systems lies in the uncertainty surrounding the types of defects the system will address and the image processing methods to be employed. Challenges such as variations in scratches, different kinds of surface wear, and the requirement for real-time processing often complicate defect detection via image processing. Typically, defect analysis can be approached using statistical, spectral, or model-based methods. Among these, machine learning a robust subset of the model-based techniques has gained significant traction in defect analysis. This study aims to identify deteriorations and defects on the surfaces of LPG cylinders used at LPG filling stations. Such deteriorations not only diminish the readability of these cylinders but also adversely affect the accuracy of cylinder reading operations. Consequently, detecting these imperfections is crucial for maintaining system accuracy. Uniquely, our method incorporates both image processing and machine learning algorithms to achieve real-time operational speeds, setting it apart from other literature. This methodology was tested on a dataset generated in an industrial setting. In our study, the dirtiness rate for LPG cylinders identified as clean was below 2%, and the average standard deviation of accuracy values for these cylinders was 0.27. In contrast, the average dirtiness rate for cylinders identified as dirty was 18%, with a higher standard deviation of 2.03.

Anahtar Kelimeler: