Background: Image-based dietary assessment provides a more intuitive approach for nutritional monitoring and health management. However, in multi-category bowl-based meals, food boundary adhesion, spatial stacking, and staple-food occlusion by upper-layer dishes still affect the accuracy of volume, weight, and nutritional composition prediction. Methods: This study proposes a nutrition prediction method for bowl-based foods by integrating semantic segmentation, multi-view three-dimensional reconstruction, and occlusion compensation. The improved DBP-FDSNet was used to extract food-category masks from top-view RGB images, while detail enhancement, boundary-assisted supervision, and spatial position encoding were incorporated to improve the segmentation quality of food boundaries and adhesion regions. The visible food surface inside the bowl was reconstructed using a bowl instance model and RGB-TSDF-based multi-view fusion, and the two-dimensional semantic results were mapped into the height-field parameter domain for category-level volume integration. For partially occluded, severely occluded, or completely invisible staple foods, a layered compensation strategy was introduced to reduce staple-food volume prediction errors and the erroneous assignment of upper-layer food volume. Food weight and whole-bowl Calories, Fat, Carbohydrate, and Protein were finally predicted using food density and a nutritional composition database. Results: DBP-FDSNet achieved a meanIntersectionoverUnion ($mIoU$) of 80.51% and a BoundaryF1 Score ($bF1$) of 85.73%. At the whole-bowl level, the MeanAbsolutePercentageError ($MAPE$) values for Calories, Fat, Carbohydrate, Protein, and total food mass were 13.23%, 18.51%, 14.18%, 13.35%, and 10.85%, respectively. Conclusions: The method improves the stability of category-level volume and nutritional composition prediction in complex bowl-based meal scenarios, providing a feasible solution for image-based dietary assessment and intelligent nutrition management.