This model classifies images of American food into 10 categories using a custom ResNet-style CNN architecture with residual connections. The hyperparameters were optimized using the ES(1+1) evolutionary strategy.

Architecture Details

Type: Custom ResNet-style CNN with 4 residual blocks
Input Size: 224×224 RGB images
Output: 10-class softmax probabilities
Parameters: ~20.6M
Framework: TensorFlow 2.x / Keras

Training Details

Parameter	Value
Optimizer	Adam
Learning Rate	0.001381 (ES-optimized)
Dropout Rate	0.256 (ES-optimized)
Label Smoothing	0.30 (ES-optimized)
Epochs	50
Batch Size	24
Data Augmentation	Random flip, rotation, zoom, contrast

Supported Classes

The model can classify the following 10 American food types:

Index	Class Name
0	chicken_wings 🍗
1	churros 🥖
2	french_fries 🍟
3	hamburger 🍔
4	hot_dog 🌭
5	ice_cream 🍦
6	macaroni_and_cheese 🧀
7	pancakes 🥞
8	pizza 🍕
9	waffles 🧇

Usage

Quick Start

python

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

# Download model
model_path = hf_hub_download(
    repo_id="HAR5HA-YELLELA/american_food_classifier",
    filename="optimized_model_best.keras"
)

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

# Class names
classes = [
    "chicken_wings", "churros", "french_fries", "hamburger", "hot_dog",
    "ice_cream", "macaroni_and_cheese", "pancakes", "pizza", "waffles"
]

# Load and preprocess image
img = Image.open("your_food_image.jpg").resize((224, 224))
img_array = np.array(img) / 255.0
img_batch = np.expand_dims(img_array, axis=0)

# Predict
predictions = model.predict(img_batch)[0]
predicted_class = classes[np.argmax(predictions)]
confidence = np.max(predictions)

print(f"Prediction: {predicted_class}")
print(f"Confidence: {confidence:.2%}")

Inference from URL

python

import requests
from io import BytesIO

def predict_from_url(url, model, classes, threshold=0.6):
    response = requests.get(url)
    img = Image.open(BytesIO(response.content)).resize((224, 224))
    img_array = np.array(img) / 255.0
    img_batch = np.expand_dims(img_array, axis=0)
    
    predictions = model.predict(img_batch, verbose=0)[0]
    top_idx = np.argmax(predictions)
    confidence = predictions[top_idx]
    
    if confidence >= threshold:
        return classes[top_idx], confidence
    else:
        return "out_of_scope", confidence

# Example
result, conf = predict_from_url("https://example.com/pizza.jpg", model, classes)
print(f"{result}: {conf:.2%}")

Performance

Metric	Value
Test Accuracy	79.80%
Validation Accuracy	79.15%
Improvement over Baseline	+8.13%

Comparison with Other Methods

Method	Test Accuracy	Improvement
Baseline CNN	73.80%	-
Keras Tuner (Hyperband)	79.27%	+7.41%
ES(1+1) (This Model)	79.80%	+8.13%

Out-of-Scope Detection

The model includes confidence-based out-of-scope detection. Images with prediction confidence below 60% are flagged as potentially out-of-scope (not one of the 10 trained food categories).

python

threshold = 0.6
if confidence < threshold:
    print("⚠️ Image may be out of scope")

Limitations

Only trained on 10 American food categories
May not generalize well to non-American cuisines
Performance may degrade on heavily processed or stylized images
Best results with centered, well-lit food photographs

Training Data

Trained on a subset of the Food-101 dataset:

Total Images: 10,000 (1,000 per class)
Train/Val/Test Split: 70% / 15% / 15%
Image Size: 224×224 pixels

Citation

If you use this model, please cite:

bibtex

@misc{yellela2025foodclassifier,
  author = {Yellela, V. Harsha Vardhan},
  title = {American Food Image Classifier with ES(1+1) Optimization},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/HAR5HA-YELLELA/american_food_classifier}
}

Author

V. Harsha Vardhan Yellela
Lawrence Technological University
MCS-5993: Evolutionary Computing & Deep Learning
Fall 2025