This project implements a binary segmentation model to detect the presence of food in images. The model combines the strengths of UNET and MobileNet for efficient and accurate segmentation.
Images: 50 food-related images curated for this project.
Annotations: Binary masks were manually created using Photoshop to label areas containing food.
Backbone: MobileNet serves as the lightweight feature extractor, providing efficiency and flexibility.
Segmentation Head: UNET architecture is employed to output binary segmentation masks.
Data Preparation:
Images were manually masked to distinguish food regions (foreground) from non-food regions (background).
Preprocessing steps include resizing, normalization, and augmentation.
The UNET-MobileNet architecture was trained to produce binary masks indicating the presence of food.
Performance metrics such as accuracy, IoU (Intersection over Union), and Dice Coefficient were used to validate the model.
Binary Segmentation: Detects whether food is present in an image with pixel-level accuracy.
Manual Annotations: High-quality ground truth masks created with Photoshop.
Lightweight Design: Utilizes MobileNet for efficient inference on resource-constrained devices.
Useful for automated food identification systems.
Can serve as a foundation for further classification or calorie estimation systems.
Explore the repository to see the code, training details, and segmented results in action!