🧪 PCB Defect Detection and Classification System – Complete Project Documentation

Project Statement

The objective of this project is to develop an automated PCB defect detection and classification system using image processing and deep learning techniques.

The system integrates:

1.Template-based image subtraction for defect highlighting

2.Contour extraction for ROI detection

3.CNN-based classification for defect labeling

4.A Streamlit-based frontend enables users to upload PCB images, view annotated results, and export defect logs (CSV) and annotated images (JPG) for documentation and reporting.

Project Objectives

🔹 Detect and localize PCB defects accurately

🔹 Classify defects into six predefined categories:

           1.Missing Hole 🕳️

           2.Mouse Bite 🐭

           3.Open Circuit ⚡

           4.Short 🔗

           5.Spur 🌱

           6.Spurious Copper 🟦

🔹 Train a robust CNN model (EfficientNet) with high accuracy

🔹 Develop a user-friendly frontend for image upload and visualization

🔹 Implement a modular backend for image inference and processing

🔹 Enable export of annotated images, defect logs, and reports

Dataset

Source:

 DeepPCB dataset

Structure:

  Paired template and test images, with annotations for defects

Processing:

          1.Image alignment using ORB + RANSAC

          2.Image subtraction for defect highlighting

          3.Thresholding with Otsu’s method

          4.ROI extraction for CNN input

Folder Structure:

PCB_DATASET/ ├── train/ │ ├── missing_hole/ │ ├── mouse_bite/ │ ├── open_circuit/ │ ├── short/ │ ├── spur/ │ └── spurious_copper/ ├── val/ │ ├── missing_hole/ │ ├── mouse_bite/ │ ├── open_circuit/ │ ├── short/ │ ├── spur/ │ └── spurious_copper/ ├── processed/ │ ├── aligned_images/ │ ├── difference_masks/ │ ├── thresholded_masks/ │ ├── defect_rois/ │ └── samples/ └── docs/

Methodology

1️⃣ Image Preprocessing

   Convert images to grayscale and normalize

   Align test images with template using ORB feature matching + RANSAC

   Apply Gaussian blur to reduce noise

   Perform image subtraction to highlight defect regions

2️⃣ Contour Detection & ROI Extraction

  Detect defect contours using OpenCV

  Crop Regions of Interest (ROI) for classification

  Label and save ROIs for training the CNN

3️⃣ CNN-based Classification

   Backbone: EfficientNetB0 (pretrained on ImageNet)

   Input: 128x128 pixels RGB

   Loss: Categorical Cross-Entropy

   Optimizer: Adam

   Training:

        1.Freeze EfficientNet base

        2.Train custom classification head

        3.Runtime data augmentation (rotation, flips, zooms)

  Output: Defect type + confidence score

4️⃣ Backend Pipeline

Modular Python functions handle:

Alignment and subtraction

ROI extraction

Model inference

Returns annotated images and defect logs

5️⃣ Frontend UI

Built with Streamlit

Features:

Image upload for template & test PCBs

Annotated visualization of defects

Adjustable confidence threshold

Download annotated images & CSV logs

Project Milestones

Milestone 1: Dataset Preparation & CV Pipeline

Dataset alignment, subtraction, thresholding, ROI extraction

Deliverables: Processed images, defect masks, ROI crops

Metrics: Alignment accuracy 100%, defect extraction 3–12 per image

Milestone 2: Model Training & Evaluation

CNN training with EfficientNet, validation & confusion matrix

Deliverables: Trained model, accuracy metrics (96% validation)

Metrics: High precision & recall, minimal misclassification

Milestone 3: Web App & System Integration

Streamlit frontend + modular backend pipeline

Deliverables: Interactive app with defect visualization

Metrics: Responsive UI, accurate defect detection, export-ready

Milestone 4: Finalization & Delivery

Export results, optimize processing, final documentation

Deliverables: Final web app, annotated images, CSV logs, PDF report

Metrics: Fully functional system, ready for demonstration

Evaluation Metrics

Metric	Description	Target
Detection Accuracy ✅	Correctly detected defects	≥95%
Classification Accuracy 🎯	Correct defect type prediction	≥95%
ROI Precision 📐	Bounding box coverage	High
Processing Time ⏱️	Time per image	≤3s
Export Quality 💾	Correct image & CSV generation	100%

Tech Stack

Area	Tools / Libraries
Image Processing 🖼️	OpenCV, Numpy
Deep Learning 🤖	TensorFlow, Keras, PyTorch
Dataset 📂	DeepPCB
Frontend 🌐	Streamlit
Backend 🛠️	Python, Modular Functions
Evaluation 📊	Accuracy, Confusion Matrix
Export 📦	CSV, Annotated Image, PDF

Project Outputs

1.Annotated PCB images with bounding boxes

2.CSV logs of defects (type & confidence)

3.Visual reports and PDF export (optional)

4.Trained EfficientNet CNN model

5.Streamlit web application for inspection

Usage Guide

1.Mount Google Drive (for Colab):

from google.colab import drive

drive.mount('/content/drive')

2.Install dependencies:

!pip install streamlit opencv-python-headless tensorflow pillow pandas

3.Run Streamlit app:

!streamlit run /content/drive/MyDrive/Akalya_PCB/milestone-4/app.py

4.Upload PCB image, adjust confidence, run inspection

5.View annotated output and download CSV/image

Future Enhancements

1)Real-time PCB inspection using camera input

2)Industrial deployment with higher resolution support

3)Semi-supervised learning for unannotated defects

4)Expanded defect categories and multi-board inspection

Author

Akalya S. – SASTRA University

PCB Defect Detection & Classification System – Full Project (Milestone 1–4)

Highlights

Transfer learning using EfficientNet 🧠

Patch-based scanning for high detection accuracy 🔎

Modular backend + interactive frontend 🖥️

Exportable CSV and annotated images for professional reporting 💾

Validation accuracy: 96% across six defect classes ✅

Summary:

This documentation combines:

Milestone 1 (Dataset & preprocessing)

Milestone 2 (Model training & evaluation)

Milestone 3 (Frontend & backend integration)

Milestone 4 (Finalization, export, and presentation)