inSpect Cosmetic Defects AI-Powered Surface Quality Control

In-line detection of visual imperfections delivering objective and consistent quality assessment.

Book a call

Detecting mathematically indeterminate defects

Surface inspection, particularly of complex exposed parts with intricate textures, demands a thorough visual acceptability review that is challenging to quantify mathematically. In this context, AI surpasses traditional mathematical analyses by evaluating the visual state of surfaces against defined quality thresholds and modelled quality control specialist’s experience, effectively determining their acceptability.

Rapid decisions based on extensive knowledge

Developing an AI model involves prototype testing and iterative feedback from experts who regularly evaluate product quality. This process aims to refine decision-making guidelines to ensure they are neither too lenient nor too strict. The AI model harnesses the collective expertise of professional quality controllers, making rapid decisions based on extensive knowledge.

Ensuring consistent standards 24/7

Manual inspections are disposed to the subjective evaluations of quality controllers, often leading to variability, increased waste, or customer complaints. AI technology addresses this by categorizing products within precise suitability thresholds, combined with modelled quality control expertise, ensuring objective and consistent quality evaluations and enhancing reliability while decreasing dependency on human judgment.

Increasing productivity, reducing waste

Aligned with Industry 5.0, this approach emphasizes adaptability and intuitiveness in production systems rather than mere automation. Enhanced productivity leads to cost efficiency by allowing workers to focus on higher-value tasks as certain quality control processes are streamlined. Simultaneously, AI ensures consistent and precise defect evaluation, minimizing material waste, rework, and resource inefficiencies throughout the production cycle.

Addressing mathematically
indeterminate surface issues
with AI

Quality control addresses both the functional and aesthetic characteristics of a product. We usually employ analytical methods to assess functionality; however, identifying cosmetic and visual defects on surfaces visible to end users presents more challenges. These defects only influence the product’s visual appeal, not functionality, and are therefore more difficult to test using traditional methods. The main challenge is the limitations of analytical tools that cannot consistently and accurately determine human acceptability of imperfections on diverse and non-homogeneous surfaces.

Such inspections require the use of industrial vision technology combined with advanced AI algorithms, including the deep learning technique (YOLO AI model), which has proven more effective than traditional models like classical U-Net in detecting these anomalies.

AI-based approaches are particularly valuable when defects are mathematically indeterminate or too numerous to calculate, or when the assessment is too subjective to ensure consistent decisions.

inSpect_Cosmetic_Defects_UI_defect_regions

inSpect_Cosmetic_Defects_UI_Real-Time_Analysis

Assessing defects according to location, size, and contrast

The AI model identifies surface defects such as dirt, foreign objects, burns, improper material injections, irregularities on transparent parts (further examined using backlighting), and other anomalies. It classifies these defects based on predefined tolerances for their location, size, and contrast.

This model processes input images into binary data, where each pixel identified as a defect is marked as ‘1’, and unaffected pixels are marked as ‘0’. It groups these pixels into clusters or ‘blobs’, sorted by the number of adjacent non-zero pixels. The model then isolates regions from the input image based on the clusters’ dimensions and positions, segregating the original pixel values into those marked as defects and those deemed acceptable. It calculates the average values of these two groups to determine the contrast, highlighting the visibility of defects.

The model then uses a template that sorts each product into two categories based on the surface type visible after assembly, so called A, B, or C class surfaces: visible and non-visible. The acceptability criteria, including defect size and contrast limits, vary a lot between categories depending on their importance to the finished products. Parameters within established thresholds can be adjusted by an operator to best determine the quality of a specific part or lot.

AI model training using real-time
production images

The model requires training with images from live production, and the more images it processes, the better its stability and sensitivity. Each image is carefully inspected, with defects marked at the pixel level to create binary images.

The YOLO (You Only Look Once) model’s segmentation component is utilized for its quick and stable predictive capabilities and advanced approach to object recognition. It classifies segments of images according to a predefined defect catalogue. Defects are analysed and marked based on their size, location, and contrast, determining their acceptability.

These annotated templates are essential for training the model, which is then evaluated and fine-tuned according to customer feedback to meet specific production needs.

Consideration must be given to the investment in equipment, model training time, and the image acquisition and annotation process, which, while substantial, yield significant long-term benefits and a strong return on investment.

inSpect_Cosmetic_Defects_AI_Detected_Defects

Dsc6337

Edge computing AI processing
power in industrial settings

In industrial deep learning systems, it’s important to consider the capabilities of the available hardware and the conditions of its operating environment. Although artificial intelligence is often associated with cloud computing, with ultra-low latency and network independency requirements, industrial applications can be integrated directly into electrical cabinets within manufacturing facilities. These environments are typically characterized by high ambient temperatures, further intensified by the heat generated from computing components.

To handle high temperatures and ensure the hardware operates efficiently within short production cycles, effective active and passive cooling is essential. Industrial computer enclosures are typically passively cooled and protected to prevent dust from entering the GPU area, while electrical cabinets are actively cooled by default to prevent overheating of other components and potential performance degradation.

Higher productivity, lower waste,
and greater workplace efficiency

AI-powered quality control significantly enhances productivity, allowing workers to spend up to 70% less time at the production/assembly line, and substantially reduces waste by more than 50% by sorting products based on defined suitability criteria rather than discarding every piece with any kind of defect. This efficient defect detection and consistent quality management also decreases resource consumption—materials and energy—benefitting environmental sustainability.

Aligned with Industry 5.0, this approach emphasizes adaptability and intuitiveness in production systems rather than mere automation. AI has the potential to leverage the experience of hundreds of knowledgeable operators to perform tasks quickly and accurately, while empowering workers to maintain full control over processes and equipment.

These advancements not only cut long-term operational costs but also improve the company’s competitive edge in markets for complex products with higher value potential. Workers enjoy more intellectually challenging tasks, driving innovation, improving user experiences, and enhancing workplace efficiency and error prevention.

inSpect_Cosmetic_Defects_AI-powered_quality_control

Technical Data

Inspection cycle

0%

total cycle time impact

Analysis scope

7-70 M

parameters checked

Scan resolution

20+ MP

per scanned image

Key Technologies

Industrial vision AI

Neural networks and deep learning models excel in tasks like complex surface inspection and quality assurance. Industrial vision AI is built with tailored technologies such as U-net, ResNET, YOLO, Faster R-CNN, Mask R-CNN, and MobileNET, delivering unmatched precision and efficiency. Fast and stable AI analysis is achieved even within typical 24VDC-based production line systems.

SDK development

Detection, analysis, and control functionalities are developed across multiple software environments within an SDK. AI models are built from scratch using fully documented source code. Every development stage is controlled, and the decisions made by the models are reviewed with the option for modification. Image analysis combines proprietary tools with vision libraries, ensuring both precision and adaptability.

System communication

The edge system communicates with production lines using a variety of protocols and standards, including DIO, OPC UA, PROFINET, Ethernet/IP, and Modbus TCP. It also supports ERP and MES database integration via web services (e.g., REST). Designed for seamless integration with virtually any MES/MOM platform, the system requires minimal software modifications for connectivity.

Sensor technology

Systems support a wide range of cameras, lenses, and vision LED lighting solutions. We have expertise with all lens types, including classic, macro, bi-telecentric, endoscopic, pericentric, hypercentric, catadioptric, boroscopic, and multiview, as well as lighting setups like direct, indirect, dome, strobe, telecentric, projector, and hybrid. Beyond vision equipment, we integrate advanced sensors such as laser distance meters, 3D profilers, stereo vision systems, and X-ray.

Related Solutions

Visual Inspection

Automated, in-line AI-driven visual inspection systems for precise and consistent defect detection and quality assessment use high-resolution imaging to eliminate human error. Inspections are non-destructive, ensuring all products meet stringent standards without any damage.

Artificial Intelligence

Applying AI models in computer vision, acoustic detection, predictive maintenance, and predictive energy management enhances anomaly detection with high accuracy and real-time monitoring, identifying subtle anomalies that traditional methods might miss.

inSpect Display Assembly

In-line visual inspection technology for monitoring defects, alignment, cleanliness, and overall quality of display assemblies as well as their software functionalities. We integrate system that use advanced vision algorithms to detect, sort, and position all assembly components.

inSpect Fuel Cells

In-line visual inspection technology for high-volume roll-to-roll fuel cell production, essential for fuel cell quality assurance. It detects and analyses defects such as scratches, cracks, particles, and pattern irregularities in real-time while monitoring crucial process parameters.

inSpect CAT / DPF

In-line vision inspection technology, using high-resolution line scan cameras, backlighting, and advanced image processing algorithms to identify cell blockages in catalytic converters and optical leaks in DPFs, as well as any deformations like chips, cracks, or gauges.

inSpect Bent Pin

In-line visual inspection technology for detecting bent pins in electronic connectors. It provides high-resolution images for defect detection and precise measurement of pin heights, mitigates reflections, and accommodates variability in pin dimensions and materials.

Frequently Asked Questions

To effectively train the AI model, several thousand images are typically required. These should be captured directly during real-time production to ensure data integrity and prevent contamination of the photographed products or components. The images, taken live with no impact on production timelines, are then converted into binary data. This data serves as the foundation for training the AI model, ensuring that the process enhances model accuracy and integrates seamlessly with production activities.
The AI model's parameters are adjustable only to a limited degree. It is designed to detect every deviation from a flawless surface and uses additional post-analysis parameters. Adjustments are restricted to defect size and contrast relative to the surrounding area. However, these adjustments are limited to a designated experienced operator and should be performed moderately to preserve the model’s accuracy and effectiveness.
The model is manually trained to recognize every potential anomaly, including both known and newly emerging defects. It is designed to detect any issue and assess its acceptability based on predefined criteria related to size, location, and contrast. Once the AI model is integrated into production, the only changes that can be made are adjustments to the tolerances for defect size and contrast. This ensures the model remains adaptive while maintaining its initial calibration.
Integrating the AI-powered quality control system does not generally require production interruptions. The only exception is the initial setup of the industrial vision system, which may necessitate brief pauses to ensure proper installation and calibration for capturing images. Once this system is in place, both the image capturing process and the integration of the model occur during live production without any need to stop production. Integration time of AI-powered quality control system is by default planned for down-time of the production line for regular maintenance activities. After physical implementation and establishment of communication with machine, a few cycles are usually needed to test the stability of the system.
The ROI depends on several factors, including the value of the product, the machine’s operational schedule (e.g., one shift per day or 24/7 operation), and the complexity of the quality inspection process. In some cases, our customers have achieved a return on investment in less than one year. For a detailed example, read our case study to see how AI-powered quality inspection transformed operations for a manufacturer of plastic products.
Yes, the AI model's decision-making process is transparent, as each decision can be traced back to specific parameters. The reason for rejection is also shown on the same image as analysed and saved aside. However, analysing each product or component requires examining between 7 million and 70 million parameters. Although the decision-making flow is fully accessible, thoroughly retracing it can be highly time-consuming due to the complexity and volume of the data involved.
To effectively train the AI model, several thousand images are typically required. These should be captured directly during real-time production to ensure data integrity and prevent contamination of the photographed products or components. The images, taken live with no impact on production timelines, are then converted into binary data. This data serves as the foundation for training the AI model, ensuring that the process enhances model accuracy and integrates seamlessly with production activities.
The AI model's parameters are adjustable only to a limited degree. It is designed to detect every deviation from a flawless surface and uses additional post-analysis parameters. Adjustments are restricted to defect size and contrast relative to the surrounding area. However, these adjustments are limited to a designated experienced operator and should be performed moderately to preserve the model’s accuracy and effectiveness.
The model is manually trained to recognize every potential anomaly, including both known and newly emerging defects. It is designed to detect any issue and assess its acceptability based on predefined criteria related to size, location, and contrast. Once the AI model is integrated into production, the only changes that can be made are adjustments to the tolerances for defect size and contrast. This ensures the model remains adaptive while maintaining its initial calibration.
Integrating the AI-powered quality control system does not generally require production interruptions. The only exception is the initial setup of the industrial vision system, which may necessitate brief pauses to ensure proper installation and calibration for capturing images. Once this system is in place, both the image capturing process and the integration of the model occur during live production without any need to stop production. Integration time of AI-powered quality control system is by default planned for down-time of the production line for regular maintenance activities. After physical implementation and establishment of communication with machine, a few cycles are usually needed to test the stability of the system.
The ROI depends on several factors, including the value of the product, the machine’s operational schedule (e.g., one shift per day or 24/7 operation), and the complexity of the quality inspection process. In some cases, our customers have achieved a return on investment in less than one year. For a detailed example, read our case study to see how AI-powered quality inspection transformed operations for a manufacturer of plastic products.
Yes, the AI model's decision-making process is transparent, as each decision can be traced back to specific parameters. The reason for rejection is also shown on the same image as analysed and saved aside. However, analysing each product or component requires examining between 7 million and 70 million parameters. Although the decision-making flow is fully accessible, thoroughly retracing it can be highly time-consuming due to the complexity and volume of the data involved.

Kontaktirajte nas

Vas zanima več? Pošljite nam svoje sporočilo, tako da izpolnite obrazec. Z veseljem bomo stopili v stik z vami v najhitrejšem možnem času.


    Vaše osebne podatke bo INEA d.o.o., Stegne 11, Ljubljana, na podlagi zakonitega interesa (komuniciranje s strankami in javnostjo, (točka f 1. odstavka 6. člena GDPR) obdelovala za namen odgovora na vaše vprašanje oz. poizvedbo. Vaše osebne podatke bomo hranili do zaključka naše komunikacije, nato jih bomo izbrisali. Če ste nam dali privolitev (točka a 1. odstavka 6. člena GDPR), da želite prejemati vam prilagojene e-novice in ponudbe, bomo predvidevanja o vaših zanimanjih obdelovali na podlagi vaše uporabe naših storitev (nakupi, povpraševanja), do vašega preklica. Od prejemanja e-sporočil se lahko kadarkoli odjavite s klikom na povezavo v prejetem sporočilu. Več o obdelavi osebnih podatkov si lahko preberete v Politiki zasebnosti.