Tune in to YOLO Vision 2024!
27 سبتمبر 2024
10:00 — 19:30 CET
Free hybrid event
Learn how the latest object detection models can help automate quality inspection in manufacturing.
Quality inspection is a critical task during manufacturing that ensures products meet the required quality standards. However, assessing quality using traditional inspection methods can be costly as product complexity increases.
Manufacturers are moving toward deep-learning-based inspection techniques, like object detection and semantic segmentation, to decrease inspection costs. Deep learning is a subfield of artificial intelligence (AI) that uses computer algorithms called neural networks to identify complex patterns in data. These techniques help automate the inspection workflow and reduce reliance on human inspectors by analyzing extensive datasets, including images and videos.
Due to its versatility and cost-effectiveness, AI-based quality assurance significantly boosts business profitability. Reports suggest that the manufacturing industry can gain more than USD 3 trillion from AI by 2035.
This article will discuss how deep learning methods can improve quality inspection and how Ultralytics YOLO11 can enhance inspection in multiple industries.
Quality inspection evaluates whether a product has defects, anomalies, or inconsistencies before reaching the consumer.
The process can occur during production, where the product moves through an assembly line or after production, but before the items move to the distribution line.
Often, it involves human experts performing visual assessments to see if the product deviates from or does not meet the desired design standards.
However, as quality demands increase, manufacturers are moving toward automated deep-learning approaches to achieve greater agility and scalability in their operations.
Deep learning approaches use artificial neural networks that work on the principles of a human brain. The networks are interconnected layers of neurons. Each neuron performs a mathematical calculation to analyze data, identify patterns, and generate a prediction.
In quality inspection, deep learning models include computer vision frameworks that automatically learn and extract features from product images.
Developing computer vision models requires experts to train a neural network on relevant datasets and run validations on a new dataset to check performance.
Once validated, experts can deploy these models on cameras and sensors using various deployment tools such as PyTorch, ONNX, and OpenVINO.
Vision-based quality inspection uses multiple methods to detect and localize damages, cracks, and missing items. The list below mentions four modern deep learning approaches.
Binary classification refers to the task of categorizing images into one of two classes, such as determining whether or not a defect is present in an object.
Based on visual data, a classification model outputs a binary yes/no decision. They help detect missing items. For example, a classification model can detect whether an item is missing or not in a product.
Multi-class classification is the task of categorizing images into more than two classes. It assigns each image to one of several predefined categories.
For example, a multi-class classification model may analyze a product’s image and return probabilities for multiple damage or crack types, indicating which one is most likely present.
This is useful in manufacturing where various defects, such as scratches, dents, or cracks, might require different handling procedures.
Localization refers to identifying the specific location of an object or feature within an image. It uses object detection models to predict bounding boxes or coordinates that highlight the specific region of damage.
This is useful for tasks like crack detection in buildings or industrial parts, where the precise location of a defect is necessary for targeted repairs.
For example, in infrastructure maintenance, localization models can analyze images of a concrete structure and mark the exact area where a crack is located.
Multi-class localization identifies and locates multiple defects within an image while also classifying each defect into one of several predefined categories.
It uses more advanced object detection models to determine a defect’s type and location to offer more detailed information.
For example, a multi-class localization model can analyze an image of a damaged item and indicate the type of defect, such as a scratch or crack, and the exact coordinates of the defect within the object.
Traditional inspection methods are more rigid, following user-defined rules and standards such as thresholds, pre-defined checklists, and pass/fail criteria.
For example, in rule-based vision techniques, experts define a particular product's ideal color, shape, and size. The system notifies the experts if a camera or other image-capturing device detects deviations from these standards.
Deep-learning approaches offer greater flexibility for building more complex detection systems. These approaches involve collecting and annotating extensive datasets of images of defective objects. Experts use the annotated data to train object detection models such as Ultralytics YOLO11. Once trained, they can deploy the model in cameras or sensors to capture images and identify defects in real time.
In the following section, we’ll take a look at how YOLO11 can be used for quality inspection.
You-Only-Look-Once (YOLO) is a state-of-the-art (SOTA) real-time object detection model famous for its high accuracy, adaptability, and speed. Its latest iteration is Ultralytics YOLO11, which improves the previous versions in terms of feature extraction, speed, accuracy, and adaptability.
It features a better architecture for more precise feature extraction and includes optimized training pipelines for faster processing speeds. It is more computationally efficient, with 22% fewer parameters and higher accuracy scores than its predecessors.
Due to its versatility, YOLO11 can help improve quality inspection workflows in multiple domains. It can help detect anomalies, damages, cracks, missing items, and packaging errors in products through performing tasks such as object detection and segmentation.
Let’s take a look at a few ways in which computer vision models can be used within the manufacturing industry.
Computer vision models can check whether a product has all the necessary items. They can detect missing components in assembled products to ensure completeness.
In electronics manufacturing, identifying missing components, misaligned parts, or soldering issues is crucial to ensure the final product is reliable and has the right functionality.
Object detection models like YOLO11 can be trained to detect missing or misplaced components on circuit boards. It can analyze images of the boards in real-time, and identify defects such as missing resistors or capacitors. This will ensure that each unit’s assembly is correct before shipment.
Crack detection is another detection task that analyzes images or sensor data to pinpoint a crack’s location, size, and severity.
The automotive industry is one example where detecting cracks in multiple components such as gears, and brake systems is necessary to ensure they meet safety standards.
Models like YOLO11 can be trained to quickly detect defects like surface scratches or cracks in complex automotive components.
Computer vision can help detect various types of damages on a product’s surface, such as scratches, dents, and deformations using computer vision tasks.
The textile industry can significantly benefit from AI-based damage detection by using object detection and segmentation models like YOLO11. It can identify defects like tears, holes, stains, or fabric inconsistencies during the production process.
Anomaly detection refers to the task of analyzing a product’s design, structure, appearance, and size to assess whether these properties deviate from the desired standards.
In pharmaceutical manufacturing , anomaly detection is vital for ensuring the quality and safety of drug products. Manufacturers can use YOLO11 to detect irregularities such as inconsistencies in tablet shapes, sizes, discoloration, or foreign particles.
Another example of how computer vision models can be used in manufacturing is within the packaging and labeling in industries. For instance, the food and beverage industry must meet strict standards for consumer safety and compliance.
Models like YOLO11 can help detect packaging errors such as incorrect labeling, damaged packaging, or missing safety seals. It can also verify that labels have correct placements with clear barcodes or expiration dates.
This ensures that products comply with industry regulations and are ready for consumer distribution.
AI-based quality inspection frameworks are still evolving and face numerous challenges. Here are a few limitations and future research directions to take into account for these technologies.
Deep-learning-based quality inspection is experiencing exponential progress due to the constant development of different object detection models. With AI-based quality inspection, manufacturers can achieve greater scalability and flexibility than traditional approaches.
Companies can use models such as YOLO11 to automate the inspection process, taking advantage of its enhanced architecture and feature extraction capabilities resulting in better accuracy and faster speed.
You can learn more about YOLO11 and other object detection models by checking out our GitHub Repository and engaging with our vibrant community. Explore how Ultralytics is redefining manufacturing through state-of-the-art deep learning frameworks.
