Image Reconstruction Error-based Industrial Image Data Drift Detection

With the wide application of neural networkbased defect detection models in industry, the problem of model performance degradation due to concept drift has been studied. For industrial images, how to convert image data into time series data for drift detection has become a research focus. This paper presents a label-free concept drift detection method for industrial images. In the proposed method, the data employed for training the defect detection model is initially processed via a reconstruction network to extract the inherent image features. In drift detection, KolmogorovSmirnov WINdowing(KSWIN) is used to track changes in reconstruction errors to determine whether the data is drifting. Among them, the reconstruction error obtained by the reconstruction network converts the image data into the time series data. A random mask is introduced into the reconstruction network to enhance the ability of the model to capture image features. Finally, The effectiveness is validated on two datasets: gradual and abrupt drift. Experimental results demonstrate its ability to accurately detect concept drift, particularly in gradual drift scenarios.