EUVIMEDEUVIMED entdecken
Crossref

Interpretable Analysis of Mask R-CNN Models

Qingfan Lei

Advances in Engineering Technology Research · 10.03.2026 · 16 (1) · 580

DOI: 10.56028/aetr.16.1.580.2026 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Advances in Engineering Technology Research, online erschienen am 10.03.2026. Verfasst von Qingfan Lei.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The black-box nature of object detection and instance segmentation models severely limits their reliable application in high-risk scenarios such as autonomous driving and industrial quality inspection. To enhance the transparency of model decision-making, this study proposes an end-to-end interpretability analysis framework for Mask R-CNN, integrating a three-stage process of training, analysis, and visualization to systematically validate the model's decision logic. Based on the COCO 2017 dataset, this study first trained a high-performance Mask R-CNN model, focusing on the detection capabilities of key categories such as persons, cars, cats, and dogs. In the interpretability analysis phase, the Gradient-weighted Class Activation Mapping (Grad-CAM) and Randomized Input Sampling for Explanation (RISE) methods were combined to generate visual heatmaps that reveal the semantic features relied upon by the model. Experiments show that Grad-CAM excels at locating local discriminative regions (e.g., vehicle wheels, animal ears), while RISE is better at outlining the overall structure of the target, improving contour integrity in occluded scenes. Their combined analysis significantly enhances the credibility of the decision basis. Furthermore, this study developed an interactive visualization toolchain supporting real-time monitoring of training metrics, multi-modal comparison of heatmaps, and automated report generation. The average single-image analysis time of the framework was substantially reduced, significantly improving model debugging efficiency. The results indicate that the model has learned visual features highly consistent with human cognition, but limitations remain in scenarios involving small objects and complex occlusion. This work provides standardized tools and methodological support for the interpretability analysis of industrial-grade vision systems, which is of great significance for promoting the development of trustworthy artificial intelligence.

Zitieren

APA

Lei, Q. (2026). Interpretable Analysis of Mask R-CNN Models. Advances in Engineering Technology Research, 16(1), 580. https://doi.org/10.56028/aetr.16.1.580.2026

Vancouver

Lei Q. Interpretable Analysis of Mask R-CNN Models. Advances in Engineering Technology Research. 2026;16(1):580. doi:10.56028/aetr.16.1.580.2026

BibTeX
@article{Lei2026,
  author = {Lei, Qingfan},
  title = {Interpretable Analysis of Mask R-CNN Models},
  journal = {Advances in Engineering Technology Research},
  year = {2026},
  volume = {16},
  number = {1},
  pages = {580},
  doi = {10.56028/aetr.16.1.580.2026}
}
Kostenlos starten →