Printed Circuit Board Defect Detection Using Machine Learning

U ovom radu opisuju se metode računalnog vida za detekciju komponenata na tiskanim pločicama. Detaljno su obrađene metode Fast RCNN i YOLO. Za praktični dio rada odabrana je YOLO metoda zbog velike brzine detekcije. Korišten je alat Google Colabratory za pisanje Python i bash skripti, te zbog dostupnih resursa grafičkih kartica za samo treniranje modela. Korišten je i radni okvir darknet koji omogućava lakši rad sa YOLO arhitekturom. Nakon samostalnog prikupljanja podataka, model je treniran, te je vrednovan korištenjem tradicionalnih mjera. Na posljetku su prikazani rezultati detekcije te sama učinkovitost dobivenog modela.This paper describes computer vision methods for detecting printed circuit board components. Fast RCNN and YOLO methods have been studied in detail. YOLO method was chosen for the practical part of the paper due to the high detection speed. Google Colabratory was used to write Python and bash scripts, and also because of the available graphics card resources. Also darknet framework was used for easier work with YOLO architecture. After data collection, the model was trained and evaluated using traditional measures. Finally, the detection results and the efficiency of the obtained model were presented