Indian Food Recipe Generator from Images 🇮🇳🍲

A deep learning-powered application that generates Indian food recipes from food images — with English to Hindi translation and top-5 dish recommendations.

Demo Video :-

Major_Procject_Recipe_Generator.mp4

🚀 Project Overview

This project aims to simplify recipe generation for Indian cuisine enthusiasts by allowing users to upload a food image and receive:

• The generated recipe in English
• An optional Hindi translation of the recipe
• Top 5 similar dish recommendations based on the uploaded food image

✨ Key Features

• 🖼️ Image Classification of 25 popular Indian dishes
• 📝 Recipe Generation using Retrieval-Augmented Generation (RAG)
• 🌐 Recipe Translation from English to Hindi (BLEU score: 83.567)
• 🍛 Dish Recommendation System (Top 5 similar dishes)
• ⚙️ Flask backend for smooth integration

---

📊 Dataset Details

1️⃣ Image Classification Dataset

• Classes: 25 Indian dishes
• Images per class: 250–300
• Total images: 6,850
• Image resolution: 224x224 px (resized for CNN)
• Examples of classes: Biryani, Dosa, Paneer Butter Masala, Chole Bhature, Idli, etc.

Metric	Value
Classes	25
Avg images/class	~274
Total images	6,850
Image size	224x224

---

2️⃣ Recipe Dataset (For Retrieval & Generation)

• Recipes per class: ~80–100
• Total recipes: ~2,300
• Fields: Dish Name, Ingredients, Preparation Steps
• Storage format: JSON
• Embedding method: Sentence embeddings using Google Gemini API
• Vector DB: FAISS

Metric	Value
Recipes	~2,300
Avg recipes/class	~92
Fields	Name, Ingredients, Steps
Vector Store	FAISS

---

3️⃣ English-Hindi Translation Dataset

• Pairs: 10,000 English-Hindi sentence pairs
• Format: JSONL ({"source": "...", "target": "..."})
• Domain: Recipe-specific sentences (ingredients & instructions)

Metric	Value
Sentence pairs	10,000
Format	JSONL
BLEU Score (after fine-tuning)	83.567

---

🛠️ Model Architecture & Training

📷 Image Classifier

• Model: CNN (Custom Sequential model)
• Accuracy: 83.0%
• Training time: 50 epochs (~1.5 hrs on GPU)

✏️ Recipe Generator (RAG)

• Embedding: Google Gemini embeddings
• Retriever: FAISS with cosine similarity
• Generator: Retrieved recipe returned as output

🌐 Translation Model

• Base model: Helsinki-NLP Opus-MT (English-Hindi)
• Fine-tuning: 10 epochs on recipe dataset
• Final BLEU Score: 83.567

🍛 Recommender System

• Method: Cosine similarity on image embeddings
• Top-N recommendations: 5 dishes

---

⚙️ Tech Stack

• Python
• TensorFlow / Keras
• Hugging Face Transformers
• Google Gemini API
• FAISS (Facebook AI Similarity Search)
• Flask
• OpenCV
• LangChain

---

📈 Results Summary

Component	Metric	Value
Image Classifier	Accuracy	83.0%
Recipe Generator	Retrieval Accuracy	~89%
Translator	BLEU Score	83.567
Recommender	Top-5 Precision	~85%

---