This paper examines the application of graph neural networks in different image processing applications, including semantic segmentation, object detection, and image classification. The subject encompasses several architectures and methodologies that integrate graph neural networks to improve the extraction and comprehension of visual characteristics by capturing both local and global contexts. The aim is to promote ongoing innovation and practical use of image processing by identifying potential approaches to advance techniques based on graph neural networks. Convolutional neural networks (CNNs) are frequently employed in the field of computer vision, specifically for the purpose of efficient photo classification. Although CNN-based classifiers have difficulties in obtaining overall features due to the constraints of convolution kernels, they also face difficulty in reliably capturing the precise locations of objects inside the picture environment. Graph neural networks are important for improving the overall performance of global feature extraction. They achieve this by analysing the connections between different nodes to acquire complete data. This paper presents a new design for a graph neural network that effectively combines local and global attention variables in order to improve prediction accuracy. In this specific design, global characteristics are obtained using a graph convolutional neural network (GCN), whereas local characteristics are gained by a CNN block. The unique method of global self-attention pooling (GSAPool) utilises the self-attention mechanism to recreate the original graph input by inserting virtual nodes. This technique automatically allocates unique weights to individual nodes in order to produce a more representative global-attention feature.