Research on Fine-Grained Classification Scheme Based on Image Recognition

Abstract:
Fine-grained image classification is to identify the general category of objects by means of fine-grained classification, but it is extremely difficult to determine the names of objects in a more refined way. The biggest challenge is that the visual differences between different categories under the same broad category are extremely small. Therefore, the image resolution required for finer classification is high. Based on this, this paper designs a novel end-to-end image fineclassification network that can focus on the spatial features of images by considering the textual and visual features of images while introducing a spatial attention mechanism, and our approach achieves good results on both Con-text and Bottles datasets. In addition, we also introduce graph neural networks for cross-modal inference, and use adaptive graph convolutional neural networks to learn a richer set of visual features and to model a more discriminative semantic space, enabling fine-grained image classification based on scene text. This research is of great importance for the long-term development of the field.