Frequency-guided image reconstruction with multi-level and multi-scale feature fusion for industrial anomaly detection

Anomaly detection has various applications in the industrial field. However, most current studies in this field are idealized such that they can only address anomaly detection in simple scenarios. For instance, most studies assumed that all detection objects are centered and typically have a homogeneous monochromatic background, with little spatial position change across different images. Such an assumption makes it difficult to apply the research findings to complex industrial production scenarios. To address the limitations of existing idealized approaches, we propose a frequency-guided image reconstruction method with multi-level feature combination (MLFC) and multi-scale feature fusion (MSFF) for anomaly detection. The overview of the proposed method is as follows: First, we employ a mixed anomaly generation strategy to generate anomalous samples for training. Second, before feeding the images into the reconstruction network, we introduce a frequency selection module. This module assigns different frequency weights to different frequency components, to suppress the tendency toward identity mapping of the anomalies of the network. Finally, we design a reconstruction network based on U-Net by employing three MLFC modules and one MSFF module. The MLFC modules in the skip-connections adaptively fuse features from multiple encoder levels using channel attention. The MSFF module thereafter fuses the multi-scale features from the decoder to enhance the representational capability of the decoder output features. The experimental results demonstrate that even with a simple U-Net as the reconstruction network, the proposed method achieves state-of-the-art performance on the metal parts defect detection dataset. Specifically, it reaches an image-level area under the receiver operating curve (AUROC) of 94.91% and a pixel-level AUROC of 97.88%, setting a new benchmark.