Visual-Only and Vision–Language Industrial Anomaly Detection Under Benchmarks and Limited-Data Settings

Abstract

Industrial anomaly detection remains challenging in manufacturing quality inspection because defects are often subtle, spatially sensitive, and costly to annotate. These constraints have motivated a recent shift from visual-only models toward vision–language approaches that can represent anomalies more flexibly. However, it remains unclear whether this shift consistently improves detection accuracy and deployment practicality under industrial conditions. To address this gap, this paper presents an experimental evaluation review of recent methods using common dataset splits, preprocessing procedures, and evaluation metrics, while retaining method-specific backbones, input resolutions, and optimization schedules. The comparison covers standard benchmarks, limited-data benchmark settings, and the Rotated EV Metal Triangle dataset derived from real electric-vehicle manufacturing data, with further analysis across texture, spatially regular object, and pose-varying object categories. Benchmark comparisons are based on single runs and interpreted descriptively, while EV dataset results are evaluated across repeated runs. Beyond image-level and pixel-level detection performance, the evaluation also considers computational efficiency to assess practical deployment trade-offs. The results show that visual-only methods achieve the strongest peak performance on standard benchmarks, with RD++ providing a favorable accuracy–speed trade-off under its evaluated configuration. However, this advantage becomes less consistent on MPDD, under limited-data benchmark settings, and on the real EV dataset, where pose variation, reduced training availability, and more challenging test conditions affect the reliability of several methods. In these settings, vision–language methods with learnable text alignment become competitive, with RareCLIP ranking among the strongest performers.