<p>Accurate dietary monitoring is vital for understanding diet’s role in health and disease, as well as for developing personalized nutrition strategies. Traditional dietary monitoring methods often rely on subjective reporting, making them time-consuming and prone to inaccuracies. Food image recognition offers a promising alternative, but challenges such as fine-grained features, diverse ingredient-dish interactions, and variations in texture and color complicate recognition tasks. To address these challenges, this paper presents the Fine-Grained Food Network (FGFoodNet), a recognition model based on the ConvNeXt architecture. FGFoodNet is the first model designed to jointly capture ingredient-dish relationships and texture-color correlations through an Ingredient Perception Attention (FIP-Attention) module and a Feature Channel Refinement Classifier (FCR-Classifier). Both modules synergistically guide the network to attend to primary ingredient features and structurally refine color- and texture-based feature channels, significantly boosting the model’s representational capacity and recognition accuracy. The experimental results show that, under the condition of maintaining comparable model parameters and computational complexity to the baseline model, FGFoodNet achieves outstanding performance across three mainstream food recognition datasets: an accuracy of 81.32% on the FoodX-251 dataset, 83.70% on the UEC FOOD-256 dataset, and an impressive 97.05% on the THFOOD-50 dataset. These results not only significantly surpass the performance of existing state-of-the-art models but also fully validate that the model achieves remarkable performance improvement while maintaining high parameter efficiency. To validate the practical application value of FGFoodNet, we developed an Android application for dietary monitoring based on the model. Users can capture food images with their smartphones to obtain dish information and personalized dietary recommendations.</p>