Cognitive visual anomaly detection with constrained latent representations for industrial inspection robot

With the fast growth of intelligent manufacturing industry, developing advanced industrial inspection robots is becoming a research and application hotspot in the fields of both computer vision and robotics. This kind of industrial inspection robots is expected to automatically detect anomalous structures (e.g., defects, damages, rejects, etc.) from the images of the manufactured products. Generally, the existing visual anomaly detection (VAD) methods mainly focus on modeling the complex and high-dimensional distribution of normal data, while neglecting the specific visual properties of abnormal data since their frequency of occurrence is much less than that of the normal data. In this paper, inspired by the human cognition on extracting abstractly visual properties and to distinguish the anomaly patterns from the observed data, we propose a novel cognitive VAD method for industrial inspection robot. Specifically, we introduce a constrained latent space to mimic the cognitive ability of humans, where the abstraction learned from the observed normal and anomaly data are represented. We build our method based on a convolutional generative adversarial network and a denoising auto-encoder, where the adversarial learning mechanism is adopted to establish the boundary between the normal and anomaly data. In the experiment, we evaluate our method on a real-world dataset where the images are captured for the manufactured products. The comprehensive results comparing with several recent VAD methods show that the proposed method is effective to detect the anomaly images of different categories with a high accuracy.