Graph-Based Pixel Representation Using GCN for Semantic Face Segmentation

Image segmentation is widely used in different fields, marking regions of interest accordingly. The most popular architectures used in this problem are convolutional neural networks, which respectively first encoder features, then reconstruct them into a labeled image. However, such methods can have problems capturing the relationships, details and relationships of objects present in an image. In this work we propose our multilayer Graph Convolutional Network in which input images are mapped into a graph structure, which provided an opportunity to use the mechanism of aggregating information from neighbors. We conduct experiments on FASSEG Instances dataset and show that our model outperforms the classic U-Net in terms of accuracy and efficiency, achieving a higher Dice score for more categories and getting mDice = 74.17%. In addition, our proposed architecture achieved a higher Accuracy = 91.24%. One of the key strengths is the significant reduction in the number of parameters required for training, from 31 032 265 (U-Net) to 1 058 313 (our model), representing a reduction in complexity of 96.59%. So our solution opens up new possibilities for creating lightweight and efficient models in image segmentation problems, surpassing U-Net, considered the benchmark in this field.