A fine-tuned EfficientNet-B0 model for classifying food images into 8 categories. Trained using a two-stage transfer learning approach with ImageNet pre-trained weights.

Model Description

Architecture: EfficientNet-B0 + custom classification head
Task: Image Classification (8 food categories)
Framework: TensorFlow / Keras
Input: 224×224 RGB images
Pre-training: ImageNet

Classes

ID	Label
0	Baked Potato
1	Burger
2	Crispy Chicken
3	Donut
4	Fries
5	Hot Dog
6	Pizza
7	Sandwich

Training

Two-Stage Transfer Learning

Stage 1 — Feature extraction: Backbone frozen, only classification head trained (LR: 1e-3)
Stage 2 — Fine-tuning: Backbone unfrozen (BatchNorm frozen), full model trained (LR: 2e-5)

Training Configuration

Parameter	Value
Base Model	EfficientNet-B0 (ImageNet)
Image Size	224 × 224
Batch Size	8
Stage 1 Epochs	12 (EarlyStopping patience=3)
Stage 2 Epochs	6
Optimizer	Adam
Loss	Sparse Categorical Crossentropy
Dropout	0.2
Data Augmentation	RandomFlip, RandomRotation(0.05), RandomZoom(0.1), RandomContrast(0.1)

Dataset

Split	Images
Train	3,797
Validation	813
Test	814
Total	5,424

Stratified 70/15/15 split with leak-free guarantee (SHA256 dedup)
Images sourced via DuckDuckGo image search, manually cleaned

Evaluation

Overall Metrics (Test Set)

Metric	Score
Accuracy	0.9328
Precision (weighted)	0.9345
Recall (weighted)	0.9328
F1 (weighted)	0.9331
F1 (macro)	0.9301

Per-Class Performance

Class	Precision	Recall	F1-score	Support
Baked Potato	0.907	0.898	0.903	98
Burger	0.934	0.904	0.919	94
Crispy Chicken	0.860	0.968	0.911	95
Donut	0.986	0.958	0.971	142
Fries	0.926	0.917	0.921	96
Hot Dog	0.957	0.957	0.957	94
Pizza	0.971	0.918	0.944	73
Sandwich	0.906	0.923	0.914	52

Available Formats

Format	File	Size	Use Case
Keras	`BestModelEfficientNetLite.keras`	16 MB	Python / TensorFlow
TFLite	`tflite/model.tflite`	4.4 MB	Mobile / Edge (dynamic range quantized)
TFLite float16	`tflite/model_float16.tflite`	7.8 MB	Mobile / Edge (float16 quantized)
TFJS	`tfjs/model.json`	15 MB	Browser / Node.js

Usage

Python (Keras)

python

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

# Load model
model = tf.keras.models.load_model(
    "BestModelEfficientNetLite.keras",
    custom_objects={"preprocess_input": tf.keras.applications.efficientnet.preprocess_input},
)

# Predict
img = Image.open("food.jpg").resize((224, 224))
x = np.expand_dims(np.array(img), axis=0).astype("float32")
probs = model.predict(x)[0]

classes = ["Baked Potato", "Burger", "Crispy Chicken", "Donut", "Fries", "Hot Dog", "Pizza", "Sandwich"]
print(f"Predicted: {classes[np.argmax(probs)]} ({probs.max():.1%})")

TFLite (Python)

python

import numpy as np
from PIL import Image
import tflite_runtime.interpreter as tflite

interpreter = tflite.Interpreter(model_path="tflite/model.tflite")
interpreter.allocate_tensors()

img = np.array(Image.open("food.jpg").resize((224, 224)), dtype=np.float32)
img = np.expand_dims(img, axis=0)

interpreter.set_tensor(interpreter.get_input_details()[0]['index'], img)
interpreter.invoke()
output = interpreter.get_tensor(interpreter.get_output_details()[0]['index'])

TFJS (JavaScript)

javascript

import * as tf from '@tensorflow/tfjs';

const model = await tf.loadGraphModel('tfjs/model.json');
const img = tf.browser.fromPixels(imageElement).resizeBilinear([224, 224]).expandDims(0).toFloat();
const predictions = model.predict(img);
const classIndex = predictions.argMax(-1).dataSync()[0];

Limitations

Trained on web-scraped images; may not generalize well to all food photography styles
Limited to 8 food categories
Best performance on clearly visible, single-item food images

License

Apache 2.0

zeyuai/efficientnet-food-classifier

EfficientNet Food Classifier