Capturing Fine-Grained Food Image Features Through Iterative Clustering and Attention Mechanisms

Food image recognition is a crucial initial step for food detection, nutritional analysis, and dietary recommendations. However, it is highly challenging due to the diverse appearance of food, which includes variations in color, texture, and presentation. We believe that effective food image recognition requires exploiting the characteristics of food images, with color playing a dominant role in feature extraction. This paper proposes a novel color-cluster attention network model for food recognition, which integrates Kohonen networks-based clustering methods with iterative feature refinement. First, the model leverages Kohonen networks-based clustering methods to effectively group similar features, which helps in capturing the complex compositions of food ingredients. Then, we apply an iterative strategy to the feature extraction module to gradually capture finer image details, and utilize a self-attention mechanism for weighting to further enhance feature representation. Extensive experiments conducted on various food image datasets, including ETH Food-101, Food11, Food new and CAFD to evaluate the effectiveness of the proposed model. The results show that the proposed model significantly improves the recognition accuracy and validate the generalization ability of the method on multiple food datasets.