Image recognition is essential in various applications, including autonomous driving and security surveillance. Traditional deep learning models, such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), have advanced image recognition by leveraging spatial and sequential representations. However, these methods struggle to capture complex object relationships, limiting their ability to fully exploit structural dependencies. To address this, recent research has explored Graph Neural Networks (GNNs) for image recognition. These methods represent images as graphs, treating image patches as nodes and establishing edges based on spatial and semantic relationships. By utilizing Graph Convolutional Networks (GCNs), GNN-based models can better model object structures and interrelations. However, high computational costs and long training times limit their practical use. This study proposes an improved GCN-based approach that enhances computational efficiency while maintaining high recognition accuracy. Specifically, we introduce a feature aggregation method that reduces redundant computations, optimizing the trade-off between accuracy and training time. Additionally, we examine the optimal placement of GCN layers in multi-layer architectures to improve performance while minimizing computational overhead. Experimental results show that the proposed method reduces training time while maintaining or exceeding the accuracy of conventional GNN approaches. These findings contribute to the development of efficient graph-based image recognition models, enabling broader adoption in real-world applications.