AI for Smart Sewer Inspection: Automatic Defect Detection from CCTV Images
Project Description
Underground sewer pipes are an essential part of urban infrastructure, but defects inside the pipes are difficult to inspect directly. In practice, CCTV cameras are commonly used to record the internal condition of sewer pipes, and engineers need to review large numbers of images and videos to identify defects such as cracks, fractures, deformation, blockage and other abnormalities.
This project explores how Artificial Intelligence (AI) and computer vision can help automate this inspection process.
The main goal is to develop an AI system that can automatically detect and identify sewer defects from individual CCTV images. Students will work with real sewer inspection data and learn how to prepare image datasets, train AI models, evaluate their performance and understand why the models succeed or fail.
Students may experiment with modern computer-vision models for image classification or object detection. Depending on their progress and interests, they may also explore vision-language models (VLMs) and other recent multimodal AI technologies to investigate whether general-purpose AI models can understand sewer defects from images and text descriptions.
The project provides a practical opportunity for students to apply modern AI techniques to a real engineering problem and to understand how AI can support the inspection and maintenance of critical urban infrastructure.
This project explores how Artificial Intelligence (AI) and computer vision can help automate this inspection process.
The main goal is to develop an AI system that can automatically detect and identify sewer defects from individual CCTV images. Students will work with real sewer inspection data and learn how to prepare image datasets, train AI models, evaluate their performance and understand why the models succeed or fail.
Students may experiment with modern computer-vision models for image classification or object detection. Depending on their progress and interests, they may also explore vision-language models (VLMs) and other recent multimodal AI technologies to investigate whether general-purpose AI models can understand sewer defects from images and text descriptions.
The project provides a practical opportunity for students to apply modern AI techniques to a real engineering problem and to understand how AI can support the inspection and maintenance of critical urban infrastructure.
Supervisor
WANG, Yu-Hsing
Quota
3
Course type
UROP1000
UROP1100
UROP2100
UROP3100
UROP3200
UROP4100
Applicant's Roles
As participants in this project, students will take part in both the AI development and experimental evaluation of the sewer inspection system. Their activities may include:
1. Data Preparation
• Extracting and organizing CCTV images from sewer inspection videos.
• Understanding common types of sewer defects.
• Preparing training, validation and testing datasets.
• Checking and improving image labels where necessary.
2. AI Model Development
• Training computer-vision models to identify defects from individual CCTV images.
• Experimenting with image-classification or object-detection approaches.
• Adjusting model settings and training strategies to improve performance.
• Comparing the performance of different AI models.
3. Modern AI Exploration
• Experimenting with modern vision models or vision-language models.
• Investigating whether AI can recognize defects using images together with natural-language descriptions.
• Comparing conventional computer-vision methods with newer multimodal AI approaches.
4. Performance Evaluation
• Evaluating model performance using suitable metrics such as precision and recall.
• Investigating which defects are easier or more difficult for AI to recognize.
• Studying false detections and missed defects.
• Visualizing and interpreting successful and unsuccessful predictions.
5. Research Demonstration
• Developing a simple prototype that demonstrates automatic sewer-defect detection.
• Presenting research findings to the project team.
• Documenting experimental results and contributing to further research or publications where appropriate.
Students are not expected to have prior experience in sewer inspection. Basic programming knowledge is helpful, but training and guidance will be provided.
1. Data Preparation
• Extracting and organizing CCTV images from sewer inspection videos.
• Understanding common types of sewer defects.
• Preparing training, validation and testing datasets.
• Checking and improving image labels where necessary.
2. AI Model Development
• Training computer-vision models to identify defects from individual CCTV images.
• Experimenting with image-classification or object-detection approaches.
• Adjusting model settings and training strategies to improve performance.
• Comparing the performance of different AI models.
3. Modern AI Exploration
• Experimenting with modern vision models or vision-language models.
• Investigating whether AI can recognize defects using images together with natural-language descriptions.
• Comparing conventional computer-vision methods with newer multimodal AI approaches.
4. Performance Evaluation
• Evaluating model performance using suitable metrics such as precision and recall.
• Investigating which defects are easier or more difficult for AI to recognize.
• Studying false detections and missed defects.
• Visualizing and interpreting successful and unsuccessful predictions.
5. Research Demonstration
• Developing a simple prototype that demonstrates automatic sewer-defect detection.
• Presenting research findings to the project team.
• Documenting experimental results and contributing to further research or publications where appropriate.
Students are not expected to have prior experience in sewer inspection. Basic programming knowledge is helpful, but training and guidance will be provided.
Applicant's Learning Objectives
By participating in this project, students will:
1. Gain Practical AI and Computer-Vision Skills
• Learn the basic workflow for developing an AI model, including data preparation, training, validation and testing.
• Gain hands-on experience with modern image-based AI models.
• Understand how computer vision can be applied to real engineering problems.
2. Work with Real Engineering Data
• Gain experience working with real CCTV inspection images rather than only standard classroom datasets.
• Understand challenges such as image quality, difficult defect appearances and imperfect labels.
• Appreciate the importance of good-quality training data in AI development.
3. Learn How to Evaluate AI Models
• Understand commonly used performance measures such as precision, recall and accuracy.
• Analyse model errors and failure cases.
• Learn how systematic experiments can be used to improve AI performance.
4. Explore Modern AI Technologies
• Gain introductory experience with modern vision models and vision-language models.
• Understand how recent multimodal AI differs from conventional task-specific computer vision.
• Explore how natural-language descriptions may be combined with images for engineering applications.
5. Develop Research and Problem-Solving Skills
• Learn how to formulate a research question and design experiments.
• Develop critical-thinking skills through analysis of successful and unsuccessful AI predictions.
• Communicate research findings through reports, presentations and demonstrations.
6. Understand AI Applications in Smart Infrastructure
• Learn how AI can support the inspection and maintenance of underground infrastructure.
• Understand the opportunities and limitations of applying AI to real engineering problems.
Students who make strong progress may further investigate more challenging topics, such as distinguishing visually similar defects, improving models with additional training data, or exploring information from nearby video frames. These will be optional extensions rather than core requirements of the project.
1. Gain Practical AI and Computer-Vision Skills
• Learn the basic workflow for developing an AI model, including data preparation, training, validation and testing.
• Gain hands-on experience with modern image-based AI models.
• Understand how computer vision can be applied to real engineering problems.
2. Work with Real Engineering Data
• Gain experience working with real CCTV inspection images rather than only standard classroom datasets.
• Understand challenges such as image quality, difficult defect appearances and imperfect labels.
• Appreciate the importance of good-quality training data in AI development.
3. Learn How to Evaluate AI Models
• Understand commonly used performance measures such as precision, recall and accuracy.
• Analyse model errors and failure cases.
• Learn how systematic experiments can be used to improve AI performance.
4. Explore Modern AI Technologies
• Gain introductory experience with modern vision models and vision-language models.
• Understand how recent multimodal AI differs from conventional task-specific computer vision.
• Explore how natural-language descriptions may be combined with images for engineering applications.
5. Develop Research and Problem-Solving Skills
• Learn how to formulate a research question and design experiments.
• Develop critical-thinking skills through analysis of successful and unsuccessful AI predictions.
• Communicate research findings through reports, presentations and demonstrations.
6. Understand AI Applications in Smart Infrastructure
• Learn how AI can support the inspection and maintenance of underground infrastructure.
• Understand the opportunities and limitations of applying AI to real engineering problems.
Students who make strong progress may further investigate more challenging topics, such as distinguishing visually similar defects, improving models with additional training data, or exploring information from nearby video frames. These will be optional extensions rather than core requirements of the project.
Complexity of the project
Challenging