Classification of Electronics Components using Machine Learning

Ng, Chong Heng (2020) Classification of Electronics Components using Machine Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Ng Chong Heng.pdf Restricted to Registered users only Download (5MB)

Abstract

A supervised machine learning is implemented to detect and classify various basic electronics components, which can help in recognize electronics component without using a complete set of electronic components libraries. With the help of TensorFlow object detection API, object detection and classification model can be trained to perform detections and multiclass classifications of electronic components using a pre-trained faster R-CNN inceptionV2 model with deep learning neural network. In this project, the targeted electronic components are capacitors, integrated circuits (IC), and resistors. Dataset is acquired in a controlled environment and also from Google image. Since the total amount of data collected is relatively low, data augmentation techniques such as transformation (rotation), the addition of random noise, and manipulation of image properties are implemented to increase the number of data while preventing overfitting. A precision and recall curve with area under curve (AUC) method is used to measure the average precision (AP) of the detections. Based on the results of the detections, the combined model which is trained using images captured from a controlled environment and obtained from Google can perform better compared to the specialized model which is trained using images captured from controlled environment only. The combined model is able to achieve a mean average precision, mAP of 85.61% with intersection over union (IOU) threshold set to 0.9 in detecting the testing data consists of images captured and Google images. In order to test the ability of generalization of the model, a new dataset is acquired where the component captured isn’t used for training the model. With the IOU set to 0.75, single-class classification using the other dataset for testing is able to achieved mAP of 61.50%. In the detections of multiple targeted electronic components in an image using the other dataset, an accuracy of 53.71% mAP with 0.75 IOU threshold is achieved using the combined model. In order to evaluate the performance of the model with the influence of false-positive components, random data consists of the targeted component and other types of components is captured and used for testing the combined model. An accuracy of 35.41% mAP with 0.75 IOU threshold is achieved. Therefore, the performance of the model is satisfied in detecting and classifying electronic components.