Automatic image recognition

Take a look!


In the claims process of the future, software will be able to use photos to recognize which auto parts have been damaged in an accident and whether a replacement is necessary. Until then, there is one task on the agenda for computers: Learn, learn, and learn again.


Multidimensional matrices, convolutional layers, fully connected layers, bounding boxes, and neural networks – if one asks Stephen Seiler, Senior Data Scientist in the Research and Development division of Control€xpert how exactly automatic image recognition works, his answers are always highly detailed. As an image recognition mastermind, the Data Scientist is totally in his element at Control€xpert. Fortunately, he knows how to explain the overall picture so that an interested layperson would not be overwhelmed by technical jargon. "Facebook, Google, Apple – all of them are working on applications such as facial recognition. These same methods and algorithms can also be used by Control€xpert on images of cars", says Kai Siersleben, Managing Director at Control€xpert, summarizing the topic. Image recognition is in fact one of the most active fields of research worldwide. "Right now everybody is doing research in this field", says Stephen Seiler. "The community is incredibly active and works closely together. The mutual exchange of knowledge benefits everyone." Due to the very complex and complicated nature of the field, it would not be possible in any other way. Hence, reading research articles and combing through blogs, tutorials, and online workshops provides the foundation for his work. The research and development team at Control€xpert has, among other things, the goal of training the software using vehicle parts and having it be able to automatically recognize them in images. Over the past few months, the team has made massive progress: A number of auto body parts can already be recognized with great accuracy. 


Success through deep learning


Just being able to make it this far required a lot of preparatory work and computing capacity. Significantly more computing power than before is required for current artificial neural networks (keyword: deep learning). These models benefit the most from parallel computing units such as the Graphics Processing Unit (GPU) used for graphics cards. We are now using multiple graphics cards in parallel in order to further accelerate model development. 
For the next step, it was necessary to obtain vehicle image material. Once these images have been collected, they are first labeled in pre-processing, meaning all visible and relevant information is repackaged into what are called label files. Once a sufficiently large dataset of labeled images exists, then the actual "learning" of the models for the computer begins. A suitable training environment is set up, and a neural network is implemented, and then it is unleashed: Deep learning begins. The duration of training ranges from a few hours to several days. Image recognition is a complex task with high computing demands. Therefore, an iterative approach is used to increase accuracy with each repetition. After the training, the actual evaluation begins, meaning: Which images has the computer correctly recognized? Which replacement parts does it now recognize better? Which predictions are not yet correct? This evaluation is also performed by computer programs. With a constantly growing image database, the computer executes learning process after learning process, thereby making its predictions more and more precise. The models are then trained until they no longer exhibit any increases in accuracy.


Faster claims calculations through image recognition


Particularly for claims auditing, automatic image recognition brings with it enormous advantages. The present claim calculation can be compared with the submitted images and be validated. Image recognition therefore directly supports the vehicle experts at Control€xpert, increasing the quality of the auditing process. This facilitates the creation of claim prognoses for EasyClaim – but much more than this is possible.