PCB komponentide tuvastamine

Abstract

The mission of this paper is to create machine learning based models for detecting the orientation of printed circuit board (PCB) components. To achieve the desired goal, during the course of this work it was necessary to create a photo-bank of different kinds of annotated PCB-s, two models, one based on Faster-RCNN with a ResNeXt backbone and the other based on YOLO. A few validation and testing programs were created to compare the upsides and downsides of each model. During the development process, it was a general rule to try use open-source solutions, so that the created programs and models would be free to use for anyone who is interested. The completed project allows anyone to connect their annotated project to model training files, to use them for directional object detection.

Käesoleva töö eesmärgiks on luua masinõppe mudel trükkplaadi komponentide suundade tuvastamiseks. Selleks valmis töö käigus pildistatud ning märgendatud andmehulk erinevatest trükkplaatidest, YOLO ja ResNeXt selgrooga FasterRCNN mudelid ning testimisprogrammid, et määrata, kumb on selliseks rakenduseks sobivam, ning anda ülevaade kummagi meetodi plussidest ja miinustest. Loodud lahenduse leidmise käigus prooviti peamiselt kasutada vabavaralisi tööriistu, et loodud lahendus oleks tasuta kasutatav kõigile huvilistele. Lahendus võimaldab soovijatel ühendada oma märgendatud projekt mudelitega, et tegeleda soovitud suunatuvastusega.