Recent advancements in neural networks has enhanced image anomaly detection in different industries such as manufacturing, reduced human intervention, and boosted productivity and quality. However, there are still some challenges in using point cloud and RGB data for anomaly detection across varying defect sizes. This study explores such challenges and and highlights the importance of high-quality data for effective training. It also highlights the trade-off between detection accuracy and inference speed, where faster models tend to sacrifice performance on smaller defects due to the use of general loss functions that do not sufficiently capture fine-grained anomalies. This work suggests that future research should focus on improving the models that efficiently detect small defects while ensuring high-speed inference and bridging the gap between accuracy and efficiency for real-world industrial applications.
The segmentation of damaged areas in vehicle images is essential for practical applications such as automated insurance processing, vehicle resale evaluation, and maintenance support. Damage patterns often appear in different forms and sizes, and exhibit irregular shapes, which makes consistent segmentation difficult. These characteristics tend to occur simultaneously in real-world environments, resulting in conditions referred to as multiple multi-scale jagged masks. To address this problem, this study designed a model to separate overlapping regions, extract features across multiple spatial resolutions, and enhance uncertain boundaries through attention-guided refinement. The proposed model was evaluated on a large-scale public dataset of annotated vehicle surface damage images using four standard metrics: intersection over union, F1-Score, precision, and recall. Comparative experiments with damage-specific and general-purpose segmentation models revealed that the proposed model achieved the highest performance across all metrics and maintained stability across repeated trials.