Indonesian Free Nutritious Meal Food Quality Inspection System: Deep Learning Application Based on CNN

Indonesian developer putrii25 released the food quality inspection system (Deteksi-Kualitas-Makanan-MBG) for Indonesia's 'Free Nutritious Meal Program' (MBG) on GitHub on June 15, 2026. This system uses Convolutional Neural Network (CNN) technology to automatically identify food ingredient quality, addressing the pain points of low efficiency and high cost in traditional manual quality inspection, and providing scalable technical support for food safety in large-scale social welfare projects.

The 'Free Nutritious Meal Program' (MBG) is an important social welfare project of the Indonesian government that provides free nutritious meals to students, pregnant women, and the poor. Large-scale food distribution faces severe challenges: traditional manual quality inspection is inefficient and costly, making it difficult to cover all supply points. This system provides an automated and scalable solution through deep learning technology, and also serves as a reference for the digital transformation of social welfare projects in developing countries.

Why Choose CNN?

CNN has four major advantages in food quality inspection (image classification tasks):

1. Local Feature Learning: Automatically learns key features such as color and texture (the difference between fresh and spoiled ingredients is obvious);
2. Hierarchical Feature Representation: Shallow layers detect basic features, while deep layers identify abstract quality patterns;
3. Spatial Invariance: Pooling layers handle uncertainties in shooting angles/compositions;
4. Parameter Efficiency: Shared convolution kernels reduce the number of parameters, making it suitable for limited data and resource-constrained devices.

Speculation on Typical Architecture

Although the detailed architecture is not publicly available, based on common patterns, it is speculated:

1. Input layer: Standardized food images (e.g., 224x224);
2. Convolution block: Convolution + ReLU + Batch Normalization + Pooling;
3. Feature extraction layer: Increase in the number of channels, decrease in spatial dimensions;
4. Global average pooling/flatten layer: Convert feature maps into vectors;
5. Fully connected layer: Map features to quality categories;
6. Output layer: Softmax outputs category probabilities.

Image Acquisition and Preprocessing

• Multi-source input: Local upload, camera capture, batch import;
• Standardization: Adjust size, normalize pixels, color conversion;
• Data augmentation: Use rotation, flipping, etc., to expand the dataset during training;
• Quality preprocessing: Denoising, contrast enhancement, etc.

Quality Classification Model

Quality grade division:

• Freshness grading (fresh/acceptable/slightly spoiled/obviously spoiled);
• Category recognition (vegetables/fruits/meats, etc.);
• Defect detection (mold/rot/pest infestation, etc.);
• Maturity assessment (optimal edible maturity for fruits)

Web Interface and User Experience

• Simple upload process;
• Color-coded visual results (green/yellow/red);
• Confidence level indication;
• Historical records and traceability;
• Multilingual support (Indonesian/English).

Data Collection Difficulties

• Class imbalance: Far more high-quality samples than spoiled ones;
• Lighting condition differences: Lighting at different times/places affects generalization;
• Variety diversity: Large appearance differences among varieties of the same ingredient;
• Shooting device differences: Uneven imaging quality;
• Annotation subjectivity: Inconsistent quality judgment standards.

Solutions

• Data augmentation: GAN-generated synthetic samples, MixUp/CutMix;
• Transfer learning: Fine-tuning based on ImageNet pre-trained models;
• Class balance: Oversampling minority classes, undersampling majority classes, Focal Loss;
• Ensemble learning: Training multiple models for integrated prediction.

Central Kitchen Quality Inspection

As a tool for incoming inspection, it automatically screens supplier ingredients and rejects non-compliant batches.

On-site Rapid Detection

Staff use mobile devices to conduct random inspections at distribution points to promptly detect transportation/storage issues.

Supply Chain Monitoring

Accumulate data to identify high-risk suppliers, problematic ingredient types, and seasonal fluctuations, and optimize strategies.

Public Supervision and Participation

Open to beneficiaries to verify quality on their own, increasing project transparency and trust.

Technical Expansion

• Multimodal fusion: Combine odor sensors, temperature records, and near-infrared spectroscopy;
• Edge computing: Deploy to edge devices like Raspberry Pi to achieve offline detection;
• Blockchain traceability: Upload detection results to the chain for certification, building a tamper-proof traceability system;
• Continuous learning: Correct errors through user feedback and update models regularly.

Social Value

• Lower technical barriers: Use open-source frameworks and pre-trained models to reduce development workload;
• Solve practical problems: Respond to public concerns about food safety;
• Replicable and scalable: Adapt to other social welfare projects;
• Capacity building: Documents and code provide learning resources for local developers.

This project demonstrates the potential of AI in the public welfare field, contributing to food safety and public health in developing countries.