isabelagregio/recipe-ai-ingredient-classifier

Recipe AI — Ingredient Classifier

Model Description

This model classifies food ingredient images into 15 categories using MobileNetV2 and TensorFlow/Keras.

It was developed as the computer vision component of Recipe AI, an end-to-end application that identifies food ingredients from images and recommends recipes using TF-IDF and nearest-neighbor retrieval.

Architecture

• Backbone: MobileNetV2
• Framework: TensorFlow / Keras
• Task: Single-label image classification
• Input: RGB image, resized to 224 × 224 pixels
• Output: 15 classification scores

Supported Classes

The output indices correspond to these classes, in order:

CLASS_NAMES = [
    "banana", "bread", "carrot", "cheese",
    "chicken-meat", "chocolate", "egg", "flour",
    "lemon", "milk", "onion", "pineapple",
    "potato", "rice", "tomato"
]

Preprocessing

Images must be converted to RGB, resized to 224 × 224, and processed using:

tf.keras.applications.mobilenet_v2.preprocess_input

Usage

import numpy as np
import tensorflow as tf
from PIL import Image
from huggingface_hub import hf_hub_download

REPO_ID = "isabelagregio/recipe-ai-ingredient-classifier"

CLASS_NAMES = [
    "banana", "bread", "carrot", "cheese",
    "chicken-meat", "chocolate", "egg", "flour",
    "lemon", "milk", "onion", "pineapple",
    "potato", "rice", "tomato"
]

model_path = hf_hub_download(
    repo_id=REPO_ID,
    filename="mobilenet_recipe_ai_ingredients.keras"
)

model = tf.keras.models.load_model(model_path)

image = Image.open("ingredient.jpg").convert("RGB")
image = image.resize((224, 224))

x = np.asarray(image, dtype=np.float32)
x = tf.keras.applications.mobilenet_v2.preprocess_input(x)
x = np.expand_dims(x, axis=0)

scores = model.predict(x, verbose=0)[0]

if not np.isclose(scores.sum(), 1.0, atol=1e-3):
    scores = tf.nn.softmax(scores).numpy()

prediction = CLASS_NAMES[int(np.argmax(scores))]
print(prediction)

Evaluation

The original project reports a classification accuracy of 90.57%.

The exact evaluation split, sample count, and additional class-level metrics should be documented before this value is interpreted as a reproducible benchmark.

Project

• GitHub Repository