Garbage Image Classification

Transfer learning compared with a small CNN trained from scratch, on a 10-class household-waste dataset.

ResNet18 fine-tuning vs a VGG-style CNN from scratch - side-by-side training, evaluation, and per-class breakdown on 10 categories of household waste

Overview • Dataset • Getting Started • Results

---

Table of Contents

• Overview
• Dataset
• Models
• Getting Started
• Project Structure
• Results
• Author

---

Overview

This project compares two approaches to image classification on real-world garbage photos:

Transfer Learning (ResNet18) Load pretrained ImageNet weights Freeze the backbone, replace the final layer Train only the new classifier head 5 epochs, ~5k trainable parameters

CNN From Scratch VGG-style architecture built by hand 4 conv blocks (32 to 256 channels) Train everything from random init 15 epochs, ~440k parameters

Both models share the same data pipeline, class-weighted loss, and cosine LR schedule so the comparison is fair.

---

Dataset

The notebook uses Garbage Classification v2 from Kaggle.

Class	Images	Class	Images
battery	756	metal	930
biological	699	paper	1336
cardboard	1411	plastic	1597
clothes	1892	shoes	1449
glass	1736	trash	453

Total: 12,259 images across 10 classes

The classes are noticeably imbalanced (clothes has roughly 4x more images than trash). A class-weighted cross-entropy loss is used to compensate.

Download

You need a Kaggle account and an API token saved at ~/.kaggle/kaggle.json. Then either:

# Kaggle CLI
kaggle datasets download -d sumn2u/garbage-classification-v2
unzip garbage-classification-v2.zip -d data/

or uncomment the relevant cell near the top of the notebook to download automatically with opendatasets.

The notebook expects images at data/standardized_256/<class_name>/.

---

Models

ResNet18 (Transfer Learning)

The pretrained ResNet18 backbone is frozen. Only the final fully-connected layer (512 x 10) is trained.

Parameter	Value
Backbone	ResNet18 pretrained on ImageNet
Trainable params	~5,100 (head only)
Optimizer	Adam lr=1e-3
LR schedule	Cosine annealing
Epochs	5

GarbageCNN (From Scratch)

A small VGG-style network: 4 conv blocks followed by a two-layer classifier with dropout.

Parameter	Value
Architecture	4x [Conv3x3 + BatchNorm + ReLU + MaxPool]
Channel progression	3 -> 32 -> 64 -> 128 -> 256
Total params	~440,000
Optimizer	Adam lr=1e-3, weight_decay=1e-4
LR schedule	Cosine annealing
Epochs	15

---

Getting Started

Prerequisites

Tool	Version
Python	3.10+
PyTorch	2.x
torchvision	0.x
Pillow	-
matplotlib	-
numpy	-

Install all dependencies:

pip install -r requirements.txt

Running the Notebook

1. Download the dataset (see Dataset above)
2. Confirm images are at data/standardized_256/<class_name>/
3. Open garbage_classification.ipynb and run top to bottom

Training runs on CPU by default. Expect around 50-75 minutes for ResNet18 (5 epochs) and 60-120 minutes for the custom CNN (15 epochs).

---

Project Structure

garbage-classification/
├── garbage_classification.ipynb    # main notebook - data prep, training, evaluation
├── requirements.txt                # python dependencies
├── .gitignore
└── README.md

The data/ and models/ folders are gitignored. You generate them locally by downloading the dataset and running the notebook.

---

Results

Model	Trainable Params	Epochs	Test Accuracy
ResNet18 (transfer)	~5,000	5	83.4%
GarbageCNN (scratch)	~440,000	15	55.3%

The gap comes down to pretrained features. ResNet18 already knows how to find edges, textures, and shapes - the visual primitives needed to separate cardboard from glass or shoes from plastic. The custom CNN has to learn all of that from scratch on a dataset that is relatively small, with some classes having under 500 training examples.

Per-class breakdown (ResNet18)

Class	Precision	Recall	F1
clothes	0.98	0.96	0.97
shoes	0.90	0.97	0.94
biological	0.98	0.85	0.91
metal	0.56	0.92	0.70
trash	0.78	0.60	0.68

Metal and trash are the hardest classes. Metal gets confused with other hard materials, and trash has the fewest training examples so the model has less to work with.

---

Author

David Geamanu

---

License

This project is available for educational purposes.