A short story of image recognition applications for long-established businesses
What does image recognition evoke to you ? Tesla’s automatic pilot mode ? Google’s automated image organization or Facebook’s face recognition system ?
All these applications are state-of-the art image recognition applications but yet they might not be the more profitable ones. Traditional business are often considered as laggard when it comes to technologic innovation but they actually carry the most added-value applications for computer vision. From automatic quality control to predictive maintenance, deeply-rooted companies are operated by many simple but repetitive tasks than can easily be automated with computer vision. But why don’t we hear about them ?
Long-established companies are facing many challenges to adapt their operations to computer vision technology. Often handicapped by their unsexy corporate images, they don’t attract talented data scientist and fall behind to develop AI applications. For this reason, many solution-provider companies started to offer a variety of off-the-self image recognition API. But once again, this approach was not satisfying. Most APIs had a too much restricted scope and performed poorly once used in the business environment.
In response to the lack of success of these APIs, more and more image recognition API providers companies are pivoting towards custom image recognition applications and it might finally be the right approach to bring AI into traditional companies’ operations. In order to tailor each system to business needs, it appeared that a strong collaboration is required between solution providers and clients. Therefore, it is relevant to present this new approach with the spectrum of value co-creation.
Co-creation principles of real-world image recognition applications
1. Custom, the system will be
As mentioned above, custom applications proved to be more way more efficient to solve businesses’ problems. Image recognition applications are systems that take in input an image and give an information about it on output. This information can be a tag (eg : there is a dog in this image) or an object localization for instance. They are highly specific to each company and therefore need to be adapted every time.
2. Client’s image, you will use
To ensure satisfying performances, each applications should be build with customers images. By that, I mean that later on the application’s system will predict information from specific images and the model used in production should be be created with extremely similar images. I won’t go into details but keep in mind, that AI learn by examples and the more relevant the examples are, the more accurate the results will be. Be careful, some images can be qualified as personal data and has to respect personal data directives.
3. Involved, your client have to be
Unlike some others IT applications, defining requirement specifications won’t be enough to build a custom applications. Customers should be involved during the whole process in order to ensure that the final application match correctly the operations. For instance, if one company wish to automate quality control, it will need to define what tags are the best to represent the different type of defect on spare parts.
4. Labelling, your client will be in charge of
Finally, in cases where the customer is the expert, the only way to create custom systems implies to put client at work. As briefly mentioned before, to build image recognition model you need to show as many example as possible. To do so, you need to annotate every images with its corresponding tags and some tags requires an expertise only possessed by operators. For instance, there is a lot of excitements around automatic cancerous cell detection on medical images. To create an auto-diagnosis system, doctors need to teach algorithm to differentiate sane and cancerous of cells and it requires a specific annotation expertise that cannot be outsourced.
Information asymmetry has inhibited computer vision applications’s development as traditional companies have struggled to understand how it could benefit their business and AI companies to uncover potential use cases for them. Establishing co-creation relationship to build image recognition application might finally allows a faster integration of AI in traditional businesses.
Deepomatic, making vision AI accessible to every businesses
Let’s illustrate how these principles can be applied to a business model. French start-up deepomatic edits a software platform enabling businesses to build custom image recognition system. Starting from simple licence plan to more project-based sales, the start-up offers support to guide clients from use case ideation to application deployment. The relationship between them and their clients is structured around step by step meetings to define scope and tags, to collect images etc. The platform they designed helps to manage dataset and performance but also bridge deepomatic’s actions to its client’s. As it is possible to improve system’s performances over time, deepomatic designed the software as a human-in-the-loop platform : once in production, the system can still return images where the system is unconfident and client’s experts can annotate again and deploy a new version. This way, system can evolve over time to match operational changes and represent a strong example of a dynamic and customized product.
For more information about deepomatic’s platform, click here.
deepomatic’s website : https://www.deepomatic.com/
Saarijävi et al (2013), “Value co-creation: theoretical approaches and practical implications”, European Business Review
Kohtamäki, Rajala (2016), “Theory and practice of value co-creation in B2B systems”, Industrial Marketing Management