Food Nutrient Estimation Based on a Dual-directional Self-attention Aggregation Enhanced Inception Model

Accurate prediction of nutritional information from food images is a critical task in intelligent dietary assessment, with broad applications in personalized health management, nutrition tracking, and public health surveillance. In this study, we propose a novel deep neural network architecture that integrates shallow global feature extraction, attention-enhanced deep representation, and multi-scale semantic fusion within a unified framework. Specifically, our model employs the early-stages of the InceptionV3 architecture to extract low-level and global features, ensuring computational efficiency while preserving essential visual cues. A dual-branch module is then introduced to enrich semantic representation: one branch comprises dual-directional attention Transformer blocks, which fuse spatial and channel attentions, while the other branch retains the Inception modules to preserve multi-scale local patterns. The fusion of these heterogeneous features yields a more expressive and discriminative representation, which is further processed by a nutrient prediction module to estimate key nutritional attributes. Extensive experiments on Nutrition5k dataset demonstrated the effectiveness of the proposed approach. Ablation studies confirm the importance of the proposed dual-branch structure and dual-directional attention module in nutrition estimation. Compared to baseline architectures, our model consistently achieves superior performance, highlighting its potential for practical deployment in real-world dietary analysis systems.