English | 中文

This repository contains the code, saved models, evaluation artifacts, and development notes for a CNN-based American Sign Language letter classifier using the Sign Language MNIST dataset.

The project compares three models: a simple baseline CNN, a tuned baseline CNN, and an enhanced VGG-style CNN. The enhanced model achieved 99.39% test accuracy on the official Kaggle test set.

• Python
• TensorFlow / Keras
• NumPy
• pandas
• scikit-learn
• matplotlib
• seaborn
• Jupyter Notebook
• Git / GitHub

• train_and_generate_report.py: full training and evaluation pipeline.
• sign_language_cnn_assignment.ipynb: notebook version of the workflow.
• docs/development/: step-by-step notes explaining how each part of the code was built.
• data/README.md: dataset download instructions.
• figures/: generated plots for class distribution, training curves, confusion matrix, and sample predictions.
• outputs/: comparison tables, classification report, confusion matrix CSV, and trained .h5 models.

Final coursework report files are not included in this public repository. This repository focuses on the code process and experiment artifacts.

The raw Kaggle CSV files are not committed. Download the Sign Language MNIST dataset from Kaggle:

https://www.kaggle.com/datasets/datamunge/sign-language-mnist

Then place these two files in data/:

data/sign_mnist_train.csv
data/sign_mnist_test.csv

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

TensorFlow is usually easier to install with Python 3.10, 3.11, or 3.12.

python train_and_generate_report.py

The script loads the dataset, preprocesses the images, trains the three CNN experiments, saves the trained models, and writes evaluation outputs into outputs/ and figures/.

The enhanced VGG-style CNN was the best model in this experiment:

• Validation accuracy: 100.00%
• Test accuracy: 99.39%
• Macro F1-score: 0.9939
• Weighted F1-score: 0.9939

The full comparison is saved in outputs/model_comparison.csv.

The coding process is documented in separate files instead of one long document:

• docs/development/01_project_setup.md
• docs/development/02_data_pipeline.md
• docs/development/03_baseline_cnn.md
• docs/development/04_enhanced_vgg_model.md
• docs/development/05_training_experiments.md
• docs/development/06_evaluation_outputs.md
• docs/development/07_saved_models.md

本仓库包含一个基于 CNN 的美国手语字母识别项目代码、已训练模型、评估结果和代码制作流程说明。数据集使用 Kaggle 的 Sign Language MNIST。

项目对比了三个模型：基础 CNN、调参后的基础 CNN、增强版 VGG-style CNN。最终增强模型在官方 Kaggle test set 上达到 99.39% test accuracy。

• Python
• TensorFlow / Keras
• NumPy
• pandas
• scikit-learn
• matplotlib
• seaborn
• Jupyter Notebook
• Git / GitHub

• train_and_generate_report.py：完整训练和评估流程。
• sign_language_cnn_assignment.ipynb：Notebook 版本的实验流程。
• docs/development/：分阶段说明代码是怎么一步步写出来的。
• data/README.md：数据集下载说明。
• figures/：类别分布、训练曲线、混淆矩阵和预测样例图。
• outputs/：模型对比表、classification report、confusion matrix CSV 和训练好的 .h5 模型。

本公开仓库不包含最终 coursework report 文件。这里主要展示代码流程、实验过程和可复现产物。

原始 Kaggle CSV 文件没有提交到仓库。请从 Kaggle 下载 Sign Language MNIST：

https://www.kaggle.com/datasets/datamunge/sign-language-mnist

下载后，把下面两个文件放到 data/：

data/sign_mnist_train.csv
data/sign_mnist_test.csv

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

TensorFlow 通常在 Python 3.10、3.11 或 3.12 上更容易安装。

python train_and_generate_report.py

脚本会读取数据、完成预处理、训练三个 CNN 实验、保存模型，并把评估结果写入 outputs/ 和 figures/。

本项目中表现最好的模型是增强版 VGG-style CNN：

• Validation accuracy: 100.00%
• Test accuracy: 99.39%
• Macro F1-score: 0.9939
• Weighted F1-score: 0.9939

完整对比结果保存在 outputs/model_comparison.csv。

代码制作流程没有写成一个总文档，而是拆成多个阶段文件：

• docs/development/01_project_setup.md
• docs/development/02_data_pipeline.md
• docs/development/03_baseline_cnn.md
• docs/development/04_enhanced_vgg_model.md
• docs/development/05_training_experiments.md
• docs/development/06_evaluation_outputs.md
• docs/development/07_saved_models.md