With the help of an intuitive web app, companies can improve their own AI models with interactive machine learning. This way, the data remains within the company. Step-by-step retraining saves resources and also makes it possible for smaller companies to use AI.
In many areas where image processing solutions are implemented, AI systems are now standard. In this case, AI refers to machine learning (ML), a sub-area of AI. Especially in machine vision, ML methods have become commonplace and are increasingly finding their way into industry.
The problem is often that these ML systems are data-hungry, and they need to be “shown” new data in order to effectively solve the company's problems. If, for example, a company currently wants to use ML to recognize new or different objects, data has to be laboriously generated and pre-processed by humans. ML experts then retrain the models until they are ready for use again. This process is extremely time-consuming and can make it difficult to use the latest ML processes, especially for small and medium-sized companies, which often have a constantly-changing product portfolio or small batch sizes.
With the help of interactive machine learning for image processing solutions, companies can control their data and to continuously modify and improve the ML models themselves. A web app has been specially developed for this purpose, which makes it possible to feed in current data. A process expert within the company can view the data and prepare it for the ML process. The intuitive interface already shows how well the current model recognizes the objects. In addition, a heat map can be generated, showing the focus of the AI on the current image and providing information about what the object detector is concentrating on. This process enables people to better understand the model’s obscure calculation method.
It becomes clear which objects and classes are not recognized so well. Images that the model is most uncertain about are displayed as a priority in the selection bar. The bounding boxes and classes of the objects can then be easily changed and saved. A further overview of the recognized classes reveals whether the ML model already recognizes certain classes well or not. It allows users of the application to understand the current model’s ability to recognize objects. These objects can also be specifically pre-processed by the user. Care has been taken to ensure that large amounts of data do not need to be pre-processed by humans - a small amount of data is sufficient to bring about incremental improvements.
Experts from the company are involved in the pre-processing work, as they know and understand their workpieces and relevant objects best. They are therefore the most suitable people to teach the AI the objects. In addition, the data to be pre-processed does not have to be passed on to third parties and therefore remains within the company.
However, the heart of this application is not only the pre-processing of the data, but also the post-training, which can be initiated by the company's experts on site. A new version of the model can then be examined in-house and improvements in object recognition identified. To optimize the model, further data can be fed in to retrain it. This process is repeated until a satisfactory result is achieved. The gradual improvement of the model enables employees on site to achieve their goals with the help of ML and to understand better how it works.
This can boost acceptance and knowledge of AI and overcome resistance to AI. Step-by-step retraining also ensures that only necessary data is pre-processed and that the models are retrained in a more resource-efficient manner, especially on smaller data sets. This is often particularly relevant for smaller companies, but also for companies that want to deal with the topic of AI in a more energy-efficient way.
The application also enables small companies to enter the field of machine learning in image processing and to define and solve their own problems using AI.
Fraunhofer Institute for Manufacturing Engineering and Automation IPA