Saliency-based Explainable Artificial Intelligence (XAI) for Interpreting U-Net Semantic Segmentation of Synthetic Liquid Spray Images

Deep learning models, particularly U-Net, have been widely used for image segmentation tasks in various domains such as medical, biological, autonomous driving, and industrial applications. However, the black-box nature of these models raises public concerns about the lack of transparency and interpretability of their decision-making processes. Unlike typical static objects, the segmentation of dynamic objects from synthetic fluid datasets are more complicated as the properties like position and shape change over time. This paper investigates these issues by exploring the uses of saliency-based Explainable Artificial Intelligence (XAI) techniques such as Class Activation Mapping (CAM), Gradient-weighted Class Activation Mapping (Grad-CAM), Grad-CAM++ and Segmentation Grad-CAM (Seg-Grad-CAM) to generate heatmap visualizations for providing insights into the prediction of U-net semantic segmentation. Using a dataset of synthetic liquid spray images, we systematically apply these techniques to visualize the model’s attention and highlights the properties and regions of the image that are influential in the model’s predictions for different classes such as lobe, ligaments and droplets. Our results demonstrate that while all techniques provide interpretable visualizations, Seg-Grad-CAM and Grad-CAM++ offer better localization of salient features, providing more precise explanations for U-Net’s segmentation decisions. The comparison of the results using quantitative analysis such as Pointing Game Accuracy and Mean Localization Error also indicated that Seg-Grad-CAM and Grad-CAM++ are more effective for interpreting complex segmentation models in liquid spray analysis.