top of page
  • Kaitos GmbH

Tire data recognition using AI - Part 2

We have developed a robust and reliable photo-based tire recognition technology that has the potential to make a big impact across the tire industry. This technology can be used in various businesses such as online tire retail, fleet management, tire storage e.g. wholesale and garages to optimize operations, improve tire inventory databases or enhance the customer experience in online retail.


As a brief review: The task is to reliably and quickly recognize tire data types such as

Brand and profile

The brand and the tire's "name".

Tire dimension

Load and speed index

DOT

based on a photo of the tire sidewall of a smartphone or some industrial equipment. This image may suffer from poor image quality, low contrast between text and background, large differences in text size and font, scratches, dirt making this task a real challenge in the field of computer vision. However, with state-of-the art AI techniques, these task has now moved into the realm of the feasible.


The AI architecture of KTR

Our development of the Kaitos Tire Recognition (KTR) software represents a significant success in the field of deep learning and computer vision. The underlying technique employed in the development of KTR is a highly modified convolutional recurrent neural network (CRNN) architecture, comprising of approximately 10 million parameters. CRNNs are end-to-end trainable, extremely robust, and capable of processing sequential data very well, making them perfectly suited for this task. More detailed information about the specific architecture and model design of KTR can be found in the first part of our blog series.


The dataset

In order to train our model, we needed a large dataset of tire images varying over a broad range of different tires. Hence, we gathered more than 300,000 tire images of complete as well as partial sidewalls piling up to 3.3 TB of image data. We dissected each image in several slices representing e.g. the dimension or the DOT of the tire only. Once we had our dataset, we needed to do the labeling. To do so, we developed an elaborated label tool based on our inhouse developed label pipelines incorporating

  1. automatized image classification - i.e. is a partial tire-sidewall, a full sidewall, several sidewalls or no tire at all on the image,

  2. automatized mask prediction - i.e. where on the image is the tire located,

  3. automatized and highly performant flattening of the tire,

  4. automatized prediction of the latest KTR model automatically provided in the UI

  5. and finally a user interface allowing to efficiently label images to continuously improve our different AI architectures leveraging step 1, 2 and especially 5.


Image of the label tool for labeling tire data
Image of the label tool we developed to provide the required data

In this way, we were able to reduce the time required to label tire images by more than a factor of 8 - i.e. starting from 4 minutes per tire, we eventually needed about 30 seconds to do the same job in a more accurate way. In addition, active learning reduced the amount of data required to achieve high accuracy by a factor of about 10. These were some of the key factors that enabled us to complete this large project with limited resources in the first place.


Training

We did the full training on our own GPU cluster. Our training of KTR from scratch to current performance included 100,000 slices, which is equivalent to about one month of permanent training on a GeForce RTX 3090.


Current results

KTR is a great progress in terms of recognition accuracy compared to our previous post. On a test set of 500 images with image quality comparable to the one given above, our model achieved a recognition accuracy of in average more than 95% within 1.5 seconds per inference on a NVIDIA Tesla T4 or an equivalent CPU based setup:

Data type

Accuracy

Brand

97.7%

Profile (main name)

94.0%

Width

98.5%

Ratio of height to width

98.5%

Rim diameter

98.5%

Load index

96.9%

Speed index

98.5%

DOT date

93.7%

Furthermore, as a by-product of our training set, our AI model can also be applied to partial areas of the tire sidewall - i.e. only the DOT or the dimension of the tire - and will thus also be a highly reliable and sophisticated dimension or DOT scanner. Finally, based on this huge amount of data we have collected, we will build as a by-product a large and up-to-date DOT database that contains both the DOT and the information about tire tread, brand, etc., and thus can link the two pieces of information.


Outlook

We are constantly retraining our AI model with new images, which leads to a steady improvement of the recognition rates mentioned above. Accuracy and speed make our software a unique tool for online tire sales, fleet management, tire storage, garages and many other business. Furthermore, with the above data read from the sidewall of a tire, a complete identification of the tire is possible, which can then be enriched with various data from some tire database as desired .


Our software will soon be available both as cloud-based or, if required, as a locally deployed API that can be easily integrated with various applications. We are confident that this will boost tire business in various fields, save time and money for businesses, and improve the customer experience. With our AI-based tire recognition software, the process of identifying or purchasing tires has never been easier.


If you would like to be kept up to date on the further development of KTR or are interested in receiving further information or are generally interested in consulting or development services in the field computer vision of AI, please do not hesitate to contact us - contact.

 

References/ recommended reading on the topic:

 

39 views0 comments

Recent Posts

See All

Comments


bottom of page