Neurocle: Deep Learning Revolutionises Automotive Inspection

by | Jun 8, 2023

Machine vision has been an integral part of the automotive industry for a considerable time. By incorporating deep learning into their manufacturing processes, manufacturers can uphold quality standards and meet customer expectations.

Machine vision offers advantages across various automobile components, such as wheels, airbags, engines, sunroofs, and mechanical parts. While conventional rule-based machine vision systems may not cover all inspection requirements, deep learning can effectively address these gaps in the manufacturing inspection process. Neurocle provides solutions to various challenges encountered in vision inspections within the automotive process and offers deep learning technology suitable for the entire automotive process.

Challenges: Frequent Overkill and Environmental Changes

R Company, which serves as a complete vehicle manufacturer involved in research, development, and production of cars, has built a strong reputation in the industry, and gained customer trust through excellent quality. Over the years, R Company has been pushing for the automation of inspection in the automotive production process. There was a strong demand to automate inspections for identifying component defects and assessing work conditions.

Recognising the potential of deep learning-based inspection to enhance vehicle quality, R Company employed deep learning engineers from agency two years ago to develop deep learning algorithms and models. However, they faced challenges in implementing the deep learning solution on the actual production line due to various environmental changes, such as frequent part replacements and process variations.

R Company had two main inspection projects they aimed to carry out by applying deep learning vision models on the production line. The first project focused on verifying the final assembly of bolts. Previously, the bolt assembly inspection was implemented using smart vision cameras. However, over-inspection occurred frequently, leading to a two-step inspection process. Initially, the inspection was performed using vision cameras, followed by manual inspection by workers. The division of inspection into two steps resulted in waste of time and resources.

The second project was to recognise VIN numbers and classify VINs for export and domestic uses. To perform the test, two deep learning models were required: text recognition (OCR) and classification. At that time, Company R was not able to reach the target inspection speed due to the poor performance of the OCR model. While facing obstacles in the automation of inspection, R Company discovered a solution from Neurocle at a domestic automation exhibition that guarantees consistent inspection performance in any process.

Neurocle’s Solution: Auto Deep Learning

■ Shortened inspection processes & an overkill rate of less than 0.5%

The cause of over-inspection in bolt assembly testing was inaccurate labelling. As the outsourced deep learning model development progressed, problems arose during the labelling process due to the limited understanding of image data by the external deep learning engineers. Neurocle’s auto deep learning technology provided an environment where internal industrial engineers, instead of outsourced engineers, could directly handle the entire modelling process, greatly improving the quality of image data and models.

Auto Deep Learning technology refers to model training algorithms that automatically select the appropriate architecture and training parameters to generate high-performance deep learning models. This feature enables users to create high-performance deep learning models with just one training session, without the need for coding. This algorithm is built into Neurocle’s training software, Neuro-T, enabling internal industrial engineers to independently handle inspection projects.

Also, company R leveraged Neurocle’s auto-labelling features to dramatically minimise the labelling resource. Industrial engineers can label massive amounts of image data within a day thanks to this feature. This is possible since this feature automatically labels images based on several images that were labelled by the user.

As a result, through model training based on Auto Deep Learning & Auto-Labeling technology, the target over-inspection rate of less than 0.5% was successfully achieved, and the inspection process was reduced from two stages to one.

■ Improved inspection speed to 0.6 seconds per item

Neurocle VIN number detection

To perform the VIN number recognition and classification project, it was necessary to apply a parallel connection of two respective models to the inspection. Therefore, both the performance of individual models and the model’s operating environment could potentially affect the inspection speed.

The roles of the two deep learning models were as follows. The text recognition model recognizes the text of each VIN number, and the classification model analyzes the form of the VIN to differentiate between domestic and export models.

Using the Auto Deep Learning software called Neuro-T, each deep learning model was generated. By running the models on multiple GPUs, the inspection time, which used to take about 2 seconds per item, was reduced to 0.6 seconds. Also, users could leverage the feature called ‘Inference Center’ where you can predict and evaluate the performance during the Proof of Concept (POC) stage before applying the model to the industrial site. Using this feature, users could change the models until they meet the optimal performance.


Available Automotive Inspections with Neurocle Solution

neurocle deep learning [Image] Application Use Case

Neurocle’s Auto Deep Learning Vision Inspection improves the quality of finished products when utilized in various inspection processes.

■ Surface Inspection

Neurocle’s deep learning model accurately detects genuine defects in press-formed components such as a motor can, car wheels, and hood components, where diffused reflections occur due to the characteristics of surface material. Also, customers can implement surface inspections with both the interior/exterior of the vehicle including seatbelt, airbag, valve component, motor, and battery. Identifying and classifying certain types of components or text is also available. Based on various projects with global automotive customers, Neurocle can share its expertise with customers and guide them to lead optimal vision inspection with optical system composition, image acquisition, and model generation.

■ Assembly Inspection

Assembly inspection projects are mostly composed of complex inspection processes. Rule-based inspection algorithms are unsuitable for complex assembly inspection projects and often result in frequent errors. On the other hand, deep learning-based defect detection is suitable for this type of inspection as it can flexibly identify variables in the visual appearance of images.

Neuro-T has successfully performed complex inspections that involve identifying multiple molding areas inside vehicle doors and detecting the presence of defects in each molding area with 99% accuracy. Neurocle’s deep learning models are ideal for projects that require multiple deep learning models or fast inspection speed. Additionally, they demonstrate excellent performance in sealing, welding, and VIN inspections.

Related articles

Trending Articles

Join our mailing list

Subscribe to our mailing list to receive regular updates!