近幾年人們持續地挖掘深度學習的潛力，深度學習也成功被廣泛地應用在各個領域，近期又出現了大量有關於如何提升深度學習方法可解釋性的研究。已經被證實的是，AI模型如果能提供決策背後原理和相關證據，對於獲得人類的信任至關重要。我們提出了一種方法－超像素圖神經網路影像解釋器，這個創新的方法目的是透過超像素演算法和圖神經網路為影像分類任務提供局部、易於理解的解釋。超像素圖影像解釋器透過運用超像素演算法將二維像素矩陣轉換成為超像素，收集低維度的資訊（例如邊緣、紋路或輪廓）並將其轉換成圖資料。再訓練圖神經網路來預測分類任務並獲得局部解釋，我們的演算法能偵測那些影響圖神經網路進行決策的關鍵節點，將這些關鍵節點轉換成原始影像裡的像素群，就能組合成我們的最終解釋。在數個資料集中，我們的方法獲得極高的解釋能力和最快的執行速度相較於傳統影像解釋方法，準確度方面和卷積神經網路相比也有競爭力。

Deep Learning models have demonstrated remarkable potential and found extensive applications across various domains. Recently, there has been a surge in research focusing on enhancing the interpretability of deep learning methods. It has been proven that an AI model's ability to offer the rationale and the relevant evidence behind specific decisions is essential in acquiring human trust. In this context, we present the Superpixel Graph Image Explainer (SPGIE), an innovative interpretable model designed to generate localized, understandable explanations for images by using superpixel algorithms and the Graph Neural Network (GNN) model. SPGIE operates by employing superpixel algorithms to convert a two-dimensional matrix of pixels to superpixels, integrating low-dimensional information such as edges, texture, or contour. By training with the GNN model and acquiring local explanations, our algorithm identifies critical nodes that influence the decision-making process of the GNN model. Consequently, we trace these superpixel nodes back to the original pixels to compose our final explanation. In experiments, our approach achieves notable interpretability power and significantly lower runtime compared to traditional Computer Vision Image Explainers, and demonstrates competitive or superior accuracy compared to Convolutional Neural Networks.

論文審定書 i
摘要 ii
Abstract iii
List of Figures v
List of Tables vi
1. Introduction 1
2. Background and Related Works 3
2.1 Superpixel Algorithm 3
2.2 Explainer of Computer Vision 4
2.3 Graph Neural Network 6
2.4 Graph Attention Network (GAT) 7
2.5 Dynamic Reduction Network (DRN) 8
3. Methodology 9
3.1 Superpixel Graph Image Classification 9
3.2 Importance Estimation of GNN Models 10
3.3 Converting Superpixels Back to Pixels 15
4. Experiments 15
4.1 Performance of Models 16
4.2 Quantitative Evaluation via Faithfulness of Image Recognition 18
4.3 DRN distance regularization 22
4.4 Qualitative Results 24
4.5 Discussion 29
5. Conclusion 30
6. References 31

Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., & Süsstrunk, S. (2012). SLIC Superpixels Compared to State-of-the-Art Superpixel Methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(11), 2274–2282. IEEE Transactions on Pattern Analysis and Machine Intelligence. https://doi.org/10.1109/TPAMI.2012.120
Adebayo, J., Gilmer, J., Goodfellow, I., & Kim, B. (2018, October 8). Local Explanation Methods for Deep Neural Networks Lack Sensitivity to Parameter Values. arXiv.Org. https://arxiv.org/abs/1810.03307v1
Avelar, P. H. C., Tavares, A. R., da Silveira, T. L. T., Jung, C. R., & Lamb, L. C. (2020). Superpixel Image Classification with Graph Attention Networks. 2020 33rd SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI), 203–209. https://doi.org/10.1109/SIBGRAPI51738.2020.00035
Chattopadhyay, A., Sarkar, A., Howlader, P., & Balasubramanian, V. N. (2018). Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), 839–847. https://doi.org/10.1109/WACV.2018.00097
Comaniciu, D., & Meer, P. (2002). Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5), 603–619. IEEE Transactions on Pattern Analysis and Machine Intelligence. https://doi.org/10.1109/34.1000236
Dwivedi, V. P., Joshi, C. K., Luu, A. T., Laurent, T., Bengio, Y., & Bresson, X. (2022). Benchmarking Graph Neural Networks (arXiv:2003.00982). arXiv. https://doi.org/10.48550/arXiv.2003.00982
Fey, M., Lenssen, J. E., Weichert, F., & Müller, H. (2018). SplineCNN: Fast Geometric Deep Learning with Continuous B-Spline Kernels. arXiv:1711.08920 [Cs]. http://arxiv.org/abs/1711.08920
Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., & Lempitsky, V. (2016). Domain-Adversarial Training of Neural Networks (arXiv:1505.07818). arXiv. https://doi.org/10.48550/arXiv.1505.07818
Gray, L., Klijnsma, T., & Ghosh, S. (2020). A Dynamic Reduction Network for Point Clouds. arXiv:2003.08013 [Hep-Ex]. http://arxiv.org/abs/2003.08013
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. https://doi.org/10.1038/nature14539
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324. https://doi.org/10.1109/5.726791
Meyer, F. (1992). Color image segmentation. 1992 International Conference on Image Processing and Its Applications, 303–306. https://ieeexplore.ieee.org/abstract/document/785528
Monti, F., Boscaini, D., Masci, J., Rodola, E., Svoboda, J., & Bronstein, M. M. (2017). Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs. 5115–5124. https://openaccess.thecvf.com/content_cvpr_2017/html/Monti_Geometric_Deep_Learning_CVPR_2017_paper.html
Ren & Malik. (2003). Learning a classification model for segmentation. Proceedings Ninth IEEE International Conference on Computer Vision, 10–17 vol.1. https://doi.org/10.1109/ICCV.2003.1238308
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. arXiv:1602.04938 [Cs, Stat]. http://arxiv.org/abs/1602.04938
Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization. 618–626. https://openaccess.thecvf.com/content_iccv_2017/html/Selvaraju_Grad-CAM_Visual_Explanations_ICCV_2017_paper.html
Stutz, D., Hermans, A., & Leibe, B. (2018). Superpixels: An Evaluation of the State-of-the-Art. Computer Vision and Image Understanding, 166, 1–27. https://doi.org/10.1016/j.cviu.2017.03.007
Vedaldi, A., & Soatto, S. (2008). Quick Shift and Kernel Methods for Mode Seeking. In D. Forsyth, P. Torr, & A. Zisserman (Eds.), Computer Vision – ECCV 2008 (pp. 705–718). Springer Berlin Heidelberg.
Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., & Bengio, Y. (2018). Graph Attention Networks. arXiv:1710.10903 [Cs, Stat]. http://arxiv.org/abs/1710.10903
Wang, Y., Sun, Y., Liu, Z., Sarma, S. E., Bronstein, M. M., & Solomon, J. M. (2019). Dynamic Graph CNN for Learning on Point Clouds (arXiv:1801.07829). arXiv. https://doi.org/10.48550/arXiv.1801.07829
Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., & Torralba, A. (2015). Learning Deep Features for Discriminative Localization. arXiv:1512.04150 [Cs]. http://arxiv.org/abs/1512.04150