Superpixel-Based Hierarchical Graph Convolution for Enhanced Interpretability in Semantic Segmentation

Despite the success of pixel-based black-box models in most semantic segmentation algorithms for autonomous driving technology, identifying the causes of recognition errors remains challenging. To address this issue, we propose a hierarchical graph convolutional semantic segmentation method based on superpixels. This approach employs a superpixel segmentation algorithm to convert the original image into graph-structured data with rich semantic information, which is then updated with a graph neural network that utilizes spatial relationships. Graph pooling aggregates results, while graph unpooling restores richer node information, leading to final semantic segmentation. The proposed method captures extensive node information and enhances the interpretability of semantic segmentation by demonstrating the process from pixels to superpixels, aggregated results, and ultimately semantic segmentation. Experimental results demonstrate that compared to GraphSAGE and TAGCN, our method improves the mean Intersection over Union by over 10% on the VOC2012 and Cityscapes datasets. This research provides a more interpretable and efficient approach to semantic segmentation in autonomous driving applications.