Case Study: Applying AI to Automate Quality Inspection in Manufacturing

Project Overview 

Conventional manufacturing frequently relies on manual quality checks, which can lead to production bottlenecks because of inspection times that are too slow and human error-caused discrepancies. With the goal of automating defect identification across various manufacturing lines, this case study explores the creation of an AI-powered system that makes use of computer vision, specifically YOLO object detection and segmentation models. A key component of this system is computer vision. It makes it possible for the AI to "see" and analyze product photos that are taken during manufacturing. 

What is Computer Vision?

Computer vision is an area of artificial intelligence (AI) that trains computers and systems to recognize and understand meaningful information from digital photos, videos, and other visual inputs. When it detects flaws or problems, it can then recommend solutions or act. It achieves this by using machine learning and neural networks.  

This is how it's implemented:  

• YOLO Object Detection: This model serves as the first filter, locating and identifying certain products within the image frame with computer vision algorithms. Imagine that after "looking" at the picture, the AI can identify every item on the conveyor belt. 

• YOLO Segmentation: After identifying an object, the system uses the bounding boxes that the object detection model produced to focus on products. Here, faults or flaws in the product image are identified using computer vision techniques like contour analysis and Canny edge detection. Consider an AI that concentrates on a particular product and then uses advanced image processing to draw attention to any imperfections in its surface.

The method automates fault detection, paving the path for quicker, more consistent, and ultimately more dependable quality control in production. This was achieved by integrating these computer vision techniques within YOLO models.

What is Quality Inspection?

The process that involves examining products, materials, or objects to ensure that they meet a specific standard or form as per requirements is Quality Inspection.

What is Defect Detection?

Finding defects or imperfections in a product during the manufacturing process is known as defect detection. This helps reduce waste material use and rework while assuring that only high-quality items are delivered to clients.

Project Goals and Objectives 

• Reduce the need for manual inspections: Processes for quality control can be automated to increase production flow.  

• Improve consistency and accuracy: By taking advantage of AI's capacity to learn and spot flaws more precisely than human inspectors.  

• Universal adaptability: Create a system that can be trained on different products that allow flexible application on various production lines.  

Our Challenges 

• Acquiring Training Data: It can take a lot of time and resources to create a sizable and varied collection of product photos with identified flaws.  

• Accuracy and Efficiency: Striking the best balance between a quicker, potentially less accurate model that is right for real-time production line integration and a highly accurate but computationally expensive model.  

The Process and Insight 

• Object Detection for Initial Image Analysis: To locate and identify specific products on the production line, a YOLO object detection model was trained on a dataset of product photos.  

• Defect Detection with Segmentation: The system used contours on defect masks and Canny edge detection to isolate anomalies, concentrating on products that were recognized based on the object detection findings. A YOLO segmentation model was then trained using the polygon points of these anomalies.  
  
  

Model Used for Training	yolov8m-seg.pt
classes	Background,Mask
epochs	100
batch_size	16
imgsz	640

• Training and Refinement of the Model: Using defect labels as training data, the YOLO segmentation model distinguished between defect regions (mask) and defect-free areas (background) in product photos.  

• Testing and Validation: The model's performance was assessed using measures such as mean average precision (MAP), accuracy, precision, and recall using a different validation dataset. 

The Results 

The YOLO segmentation model achieved an overall accuracy of 86.7%. While mask precision remained high at 80%, mask recall reached only 50%. This indicates that the model effectively identified defect regions (high precision) but might miss some defects (lower recall). 

Accuracy	86.7%
Mask Precision	80%
Mask Recall	50%
Min Avg Precision	55%
MAP	42%

Final Takeaway 

The bottom line: Progress but still room for improvement  

The AI system's preliminary findings are positive! It successfully and accurately (86.7%) detected product flaws. This indicates that the AI identified most flaws and learned from the training set.  

Recall is one area, though, where we still have room for improvement. The term recall describes the system's capacity to identify every flaw that exists. Although the AI can detect flaws with high precision that it has previously encountered, it may miss some that are unique or uncommon (poor recall). 

Consider this: when the AI is being trained, show it photographs of numerous phone cases that have scratches on them. In the future, it will be excellent at identifying scratches. However, since the little crack in the phone casing wasn't included in its training set, it might not identify as a flaw if it comes across one instead.  

More data is needed to solve this! We may train the AI to be more accurate in its fault identification capabilities by gathering photographs of a greater range of defects. We'll also keep improving the model itself so that it can recognize even minute abnormalities.  

To put it briefly, the preliminary findings are encouraging for a trustworthy fault detection system. With more information and continuous work, we're sure we can create a developing an effective AI system that greatly enhances manufacturing quality control processes. 

Looking Forward  

We are committed to working collaboratively with the client to refine the model, address challenges, and ultimately deliver a reliable and adaptable AI-driven quality inspection system.