YOLO-Based Deep Learning Model for Multi-Class Defect Detection in the PCB Manufacturing

İzemnur Budak; Sezen Bal; Hayriye Korkmaz

Makale detayları

YOLO-Based Deep Learning Model for Multi-Class Defect Detection in the PCB Manufacturing

Yazarlar:

İzemnur Budak | Sezen Bal | Hayriye Korkmaz

Yayıncı:

Afyon Kocatepe University

Ana Başlıklar:

Bilgisayar Görüşü ve Çoklu Ortam Hesaplama (Diğer) | Yazılım Mühendisliği (Diğer)

Dergi:

Afyon Kocatepe Üniversitesi Fen Ve Mühendislik Bilimleri Dergisi Vol.25 Issue 4, pp.816-826

Belge Tipi:

Makale

Yayın Tarihi:

4/2025

Belge Dili:

Türkçe

Document ID (Org No.):

2001354992

pISSN:

2149-3367

Anahtar kelimeler:

Özet:

 In this study, a defect detection and classification method using image processing techniques and deep learning algorithms is proposed, which aims to minimize human-induced errors and speed up the production process by automatically detecting possible defects encountered in the PCB production process. In particular, the applicability of such a deep learning-based system in the PCB production line was evaluated and a real-time simulation was performed in this direction. In this study, 1194 PCB images were used in the open-source DeepPCBdataset, which consists of images containing different types of defects. A total of 7897 defects were labeled on the images in the data set, of which 6385 80 were reserved for training and 1512 20 for testing. The training and test sets were randomly divided to provide a balanced distribution among the classes. Defects were labeled using Make Sense software and six different defect types were identified during this process. The labeling process was considered a critical step in achieving high accuracy during the training phase of the model. The model trained with the YOLOv4 algorithm had a mAP50 value of 64.1. In the model trained using YOLOv8, a more up-to-date algorithm, the mAP50 value reached 93.5. These results show that the YOLOv8 algorithm offers a viable solution with high accuracy and fewer errors in PCB production lines. On the other hand, the default standard parameters of YOLOv4 and v8 were used in these models. However, when the literature was reviewed, it was seen that optimizing these parameters had an important share in achieving higher performance.

 In this study, a defect detection and classification method using image processing techniques and deep learning algorithms is proposed, which aims to minimize human-induced errors and speed up the production process by automatically detecting possible defects encountered in the PCB production process. In particular, the applicability of such a deep learning-based system in the PCB production line was evaluated and a real-time simulation was performed in this direction. In this study, 1194 PCB images were used in the open-source DeepPCBdataset, which consists of images containing different types of defects. A total of 7897 defects were labeled on the images in the data set, of which 6385 80 were reserved for training and 1512 20 for testing. The training and test sets were randomly divided to provide a balanced distribution among the classes. Defects were labeled using Make Sense software and six different defect types were identified during this process. The labeling process was considered a critical step in achieving high accuracy during the training phase of the model. The model trained with the YOLOv4 algorithm had a mAP50 value of 64.1. In the model trained using YOLOv8, a more up-to-date algorithm, the mAP50 value reached 93.5. These results show that the YOLOv8 algorithm offers a viable solution with high accuracy and fewer errors in PCB production lines. On the other hand, the default standard parameters of YOLOv4 and v8 were used in these models. However, when the literature was reviewed, it was seen that optimizing these parameters had an important share in achieving higher performance.

 In this study, a defect detection and classification method using image processing techniques and deep learning algorithms is proposed, which aims to minimize human-induced errors and speed up the production process by automatically detecting possible defects encountered in the PCB production process. In particular, the applicability of such a deep learning-based system in the PCB production line was evaluated and a real-time simulation was performed in this direction. In this study, 1194 PCB images were used in the open-source DeepPCBdataset, which consists of images containing different types of defects. A total of 7897 defects were labeled on the images in the data set, of which 6385 80 were reserved for training and 1512 20 for testing. The training and test sets were randomly divided to provide a balanced distribution among the classes. Defects were labeled using Make Sense software and six different defect types were identified during this process. The labeling process was considered a critical step in achieving high accuracy during the training phase of the model. The model trained with the YOLOv4 algorithm had a mAP50 value of 64.1. In the model trained using YOLOv8, a more up-to-date algorithm, the mAP50 value reached 93.5. These results show that the YOLOv8 algorithm offers a viable solution with high accuracy and fewer errors in PCB production lines. On the other hand, the default standard parameters of YOLOv4 and v8 were used in these models. However, when the literature was reviewed, it was seen that optimizing these parameters had an important share in achieving higher performance.

 Bu çalışmada, PCB üretim sürecinde karşılaşılan olası kusurları otomatik olarak tespit ederek insan kaynaklı hataları en aza indirmeyi ve üretim sürecini hızlandırmayı amaçlayan; görüntü işleme tekniklerini ve derin öğrenme algoritmalarını kullanan bir kusur tespit ve sınıflandırma yöntemi önerilmiştir. Özellikle PCB üretim hattında böyle derin öğrenme tabanlı bir sistemin uygulanabilirliği değerlendirilmiş ve bu doğrultuda gerçek zamanlı bir simülasyon gerçekleştirilmiştir. Bu çalışmada farklı kusur türlerini içeren görüntülerden oluşan açık kaynaklı DeepPCB veri seti içinde yer alan 1194 PCB görüntüsü kullanılmıştır. Veri setinde yer alan görüntüler üzerinde toplam 7897 kusur etiketlenmiş ve bunlardan 6385 adedi 80 eğitim, 1512 20 adedi ise test için ayrılmıştır. Eğitim ve test setleri, sınıflar arasında dengeli bir dağılım olacak şekilde rastgele bölünmüştür. Kusurlar, MakeSense yazılımı kullanılarak etiketlenmiş ve bu süreçte altı farklı kusur türü tanımlanmıştır. Etiketleme işlemi, modelin eğitim aşamasında yüksek doğruluk elde edilmesinde kritik bir adım olarak değerlendirilmiştir. YOLOv4 algoritmasıyla eğitilen model, mAP50 değeri 64,1 olarak hesaplanmıştır. Daha güncel bir algoritma olan YOLOv8 kullanılarak eğitilen modelde ise mAP50 değeri 93,5e ulaşmıştır. Bu sonuçlar, YOLOv8 algoritmasının, PCB üretim hatlarında yüksek doğruluk oranı ve daha az hata ile uygulanabilir bir çözüm sunduğunu göstermektedir. Diğer taraftan bu modellerde, YOLOv4 ve v8in varsayılan standart parametreleri kullanılmıştır. Ancak literatür taraması yapıldığında bu parametrelerin optimize edilmesinin daha yüksek performans elde edilmesinde önemli bir payı olduğu görülmüştür.

Makale detayları

YOLO-Based Deep Learning Model for Multi-Class Defect Detection in the PCB Manufacturing

Yazarlar:

İzemnur Budak | Sezen Bal | Hayriye Korkmaz

Yayıncı:

Afyon Kocatepe University

Ana Başlıklar:

Bilgisayar Görüşü ve Çoklu Ortam Hesaplama (Diğer) | Yazılım Mühendisliği (Diğer)

Dergi:

Afyon Kocatepe Üniversitesi Fen Ve Mühendislik Bilimleri Dergisi Vol.25 Issue 4, pp.816-826

Belge Tipi:

Makale

Yayın Tarihi:

4/2025

Belge Dili:

Türkçe

Document ID (Org No.):

2001354992

pISSN:

2149-3367

Defect Detection

PCB

Image Processing

Deep Learning

Real-Time Application

Manufacturing Line AutomationKusur Tespiti

PCB

Görüntü İşleme

Derin Öğrenme

Gerçek Zamanlı Uygulama

Üretim Hattı Otomasyonu

Sayı

Makale detayları

Yazarlar: İzemnur Budak | Sezen Bal | Hayriye Korkmaz

Yayıncı: Afyon Kocatepe University

Ana Başlıklar: Bilgisayar Görüşü ve Çoklu Ortam Hesaplama (Diğer) | Yazılım Mühendisliği (Diğer)

Dergi: Afyon Kocatepe Üniversitesi Fen Ve Mühendislik Bilimleri Dergisi Vol.25 Issue 4, pp.816-826

Belge Tipi: Makale

Yayın Tarihi: 4/2025

Belge Dili: Türkçe

Document ID (Org No.): 2001354992

pISSN: 2149-3367

Sayı

Yazarlar:

İzemnur Budak | Sezen Bal | Hayriye Korkmaz

Yayıncı:

Afyon Kocatepe University

Ana Başlıklar:

Bilgisayar Görüşü ve Çoklu Ortam Hesaplama (Diğer) | Yazılım Mühendisliği (Diğer)

Dergi:

Afyon Kocatepe Üniversitesi Fen Ve Mühendislik Bilimleri Dergisi Vol.25 Issue 4, pp.816-826

Belge Tipi:

Makale

Yayın Tarihi:

4/2025

Belge Dili:

Türkçe

Document ID (Org No.):

2001354992

pISSN:

2149-3367