Calorie detection in dishes based on deep learning and 3D reconstruction

Caloric intake is a crucial factor in maintaining overall health. This work addresses the challenge of automating detection of calories in vegetables and dishes by utilizing an Intel Realsense depth camera and implementing deep learning techniques in image classification combined with 3D reconstruction to the field of food. By incorporating a CBAM mechanism before and after the DenseNet264 convolutional layer, the accuracy rate is improved to 83.74%. Furthermore, a calorie detection model with a maximum average error of 0.0429 is developed through the integration of 3D reconstruction, principal component analysis and the Monte Carlo volume algorithm. Specifically, the average errors for individual dishes range from 0.049 to 0.102, while set meal samples have errors between 0.017 and 0.075. The errors observed in this study are maintained within acceptable limits, suggesting potential applicability for robot-assisted intelligent catering services in the future.