<![CDATA[Visual inspection plays an important role in maintaining product quality in automotive manufacturing. In roof inspection, abnormalities may occur across different vehicle types and require consistent inspection of both right and left sides. This study investigates the application of a PatchCore-based anomaly detection approach for roof abnormality inspection in an automotive manufacturing environment. The proposed system utilizes two cameras to capture the right and left roof areas, while the inspection result is classified as normal or abnormal through an AI-based model. PatchCore, implemented using the Anomalib library, was selected as the anomaly detection model. The dataset was determined based on production volume and defect information, resulting in 823 normal data and 44 synthetically generated anomaly images for model development and evaluation. Image preprocessing consisted of background removal and calibration, cropping, resizing to 256 pixels, and grayscale conversion. Before model development, outlier data such as blurred images, incorrect capture positions, wrong vehicle types, hand occlusion, and lighting issues were identified and excluded from training. The model was evaluated using prediction score distributions, AUROC, and recall, followed by miss-prediction analysis and iterative improvement of the training data and preprocessing process. The analysis showed that variations in brightness, capture position, and preprocessing could affect the separation between normal and anomalous data and could cause certain anomalies to be missed when anomaly-related regions were removed during preprocessing. Based on the final evaluation, the practical detection level of this AI model reached 95% for Vehicle Model_A and 86% for Vehicle Model_B, determined using recall values. Based on the observed prediction score distributions in this study, an operational uncertainty range between 0.4 and 0.6 was defined for additional human verification. The results demonstrate that PatchCore can support roof abnormality inspection while highlighting the importance of data quality, preprocessing consistency, and iterative model analysis for practical implementation in automotive manufacturing.]]>
Copyrights © 2026
