Interactive Explainable Anomaly Detection for Industrial Settings

TL;DR

The NearCAIPI Interaction Framework, which improves AI through user interaction, is presented, and it is shown how this approach increases the system's trustworthiness and how it can integrate human feedback into an interactive process chain.

Abstract

Being able to recognise defects in industrial objects is a key element of quality assurance in production lines. Our research focuses on visual anomaly detection in RGB images. Although Convolutional Neural Networks (CNNs) achieve high accuracies in this task, end users in industrial environments receive the model's decisions without additional explanations. Therefore, it is of interest to enrich the model's outputs with further explanations to increase confidence in the model and speed up anomaly detection. In our work, we focus on (1) CNN-based classification models and (2) the further development of a model-agnostic explanation algorithm for black-box classifiers. Additionally, (3) we demonstrate how we can establish an interactive interface that allows users to further correct the model's output. We present our NearCAIPI Interaction Framework, which improves AI through user interaction, and show how this approach increases the system's trustworthiness. We also illustrate how NearCAIPI can integrate human feedback into an interactive process chain.

Figures (14)

• Figure 3: Illustration of how the human interaction pipeline works. First, an image with the highest potential for information gain is selected. For this image, the AI predicts the class and explains its decision to the human expert. We generate refutations depending on the expert's feedback and add the image and the refutations to the training dataset. If the prediction is wrong, we also expect feedback from the user regarding the nearest hit-and-miss of the image.
• Figure : (a)
• Figure : (a)
• Figure : (a) Input image: NOK
• Figure : (a)