Tire Recognition meets Machine learning v.2: Step by step approach – Frank’s case [PART 2]

While in first part about Frank’s case we explained what the problem was, modus operandi and taken steps, here we are going to describe approach conducted and what was the final output and business benefits.

Taken approach - step by step  

1. PoC – choose the scope

Building trust with a client and demonstrating the solution is usually achieved through Pilot/PoC stage. Choosing a scope as a small percentage of total unique tire number which need to be identified is a good starting point. All the necessary components of the system will be present within the Pilot, but the data acquisition/annotation phase will be much shorter and RoI might be easily calculated after it.

2. Monitoring the data acquisition and labeling the data

During the data acquisition phase ardent monitoring should be performed either to detect some algorithmic flaws, system setup flaws or careless behavior by the people putting the tires on the conveyor belt. Allowing the errors to continue will in all cases prolong the Pilot phase, which is not in anyone’s interest, but most importantly it will erode the trust with the client. Make sure you have very high level of confidence in your software and procedures before pointing fingers at employees.

Labeling the data should be performed in parallel with the data acquisition phase and is supported by custom app developed to speed up the process as much as possible using suggestions from machine learning models.

3. The Pipeline

1. First we captured the images of a tire traveling on conveyer belt and detect inner and outer radius which will allow us to dewarp the tire (transform it into a trackline).

As it can be seen for this example we haven’t captured a frame where the whole tire is present. We have developed a specific algorithm to stitch different frames to get the whole tire picture. This reduces the costs needed for camera hardware by one order of magnitude.

Based on the key points in both frames we can merge them to get the complete tire picture which we can push down the processing pipeline.

On such image text is located, and that text is used in a process of identification of tire. Using in-house built text classification and text reading (OCR) models we can identify trademark, model, dimension, load index, speed index and DOT written of tire sidewall. 

And finally, the results are compared with the existing inventory catalogue for error correction and filling out the needed data.

Detect – Classify – Improve – Deliver – Profit

It is important to present the client with clear numbers. We have matched human performance (you would be surprised by the skill level of people that are manually shuffling the tires around) in the upper 90% and allowed businesses to scale their operations. Implementing AI based solutions inside industrial environments is always a challenge. This time we successfully overcame what was thrown at us.

To conclude: Do not try this at home

Industrial AI applications always require a big upfront investment to cover the development costs and usually the resources required to do so are not available to hobbyists or individuals. And even more so, products in the industrial domain require a dedicated team of professionals focused on solving problems which pop up (inevitably) and were not predicted in planning sessions or when discussing implementation details.