Enhancing Local Context in Vision GNN using Depthwise Convolution
Year: 2026
Type: conference-paper
Abstract: Convolutional Neural Networks and Transformers have emerged as dominant architectures in the field of computer vision. However, these models face limitations in flexibly representing irregular and complex objects due to their reliance on fixed grid or sequence structures. As an alternative, Vision GNN converts images into graph structures, effectively capturing semantic relationships between nodes based on k-nearest neighbors rather than physical proximity. Nevertheless, since Vision GNN processes information at the node level, it often overlooks the physical continuity of pixels, resulting in a deficiency of local context. This limitation leads to difficulties in distinguishing between high-frequency noise and actual edges, causing the model to erroneously connect unrelated nodes based solely on morphological similarity. To address these issues, this paper proposes a novel model incorporating a Convolutional Feed-Forward Network (ConvFFN) module designed to enhance local context. The proposed model introduces depthwise convolution into the existing Feed-Forward Network, presenting a clear distinction from SegFormer, which applies this concept solely to Transformers, and WiGNet, which relies on restrictive window partitions. The proposed model integrates local inductive bias into the Vision GNN without disrupting information flow. In this hybrid architecture, fully connected layers handle channel-wise information exchange, while depthwise layers are dedicated to local feature learning. Consequently, the proposed structure effectively mitigates noise and captures local context with minimal computational overhead. Furthermore, it achieved a Top-1 accuracy of 85.63% on the CIFAR-10 dataset, demonstrating a 7.07% improvement over the baseline Vision GNN.
Institution: Kangnam University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202274546