CPE 4903: Neural Networks and Machine Learning

This project aims to train two Convolutional Neural Networks (CNNs): 'Hand Digit Recognition' and 'Cats and Dogs Classification.' The models will then be deployed on a Raspberry Pi to perform inference through PiCamera.

The Hand Digit Classifier was required to have an accuracy of $\geq 99\%$ whereas the Cats and Dogs Classification model must have an accuracy of $\geq 85\%$.

Design and Implementation

Hardware Specifications

Both models were deployed on the same device. The products and their specifications are:

• Raspberry Pi 4 - Model B - 2GB RAM - 64-bit quad-core Cortex-A72 Processor
• Raspberry Pi Camera Module - imx219

Software Specifications

The software specifications along with the Python libraries are:

• Raspbian 32-bit OS, Bookworm
• Python 3.7.12
• TensorFlow 2.4.0
• Keras Preprocessing
• OpenCV
• Pyenv for creating a local environment for Python

Training the Models

Hand-Digit Recognition

The dataset used for training the model MNIST. The dataset was split into training and testing, with 60000 images for training and 10000 for testing. The size of the input images was $(28, 28, 1)$.

The model was trained using TensorFlow and Keras on Google Colab. It's a custom model consisting of 4 convolutional layers and 3 dense layers. Max pooling was also used after every convolutional layer.

Cats and Dogs Classification

The dataset used for training consisted of images of cats and dogs. It was split into training, validation, and test subsets with a ratio of 60%-15%-25% using sklearn.model_selection.train_test_split() function. A total of 10,000 images were used, with 5000 images of cats and 5000 images of dogs. After splitting the data, there were

• 6500 images in the training set
• 1506 images in the validation set
• 1994 images in the testing set

The model was created using TensorFlow and Keras on Google Colab. Ensemble learning was used to get more accurate results. VGG-16 was used and on top of that, 2 fully connected layers were added along with a Dropout layer of $30\%$. The pre-trained model, trained on the ImageNet dataset, was used and all the initial layers were frozen. All of the images were reshaped to $(224, 224, 3)$ for better results.

Environment

Setting Up the Environment

To use the Raspberry Pi for image classification, a virtual environment with an older version of Python was used. The instructions are provided here.

Other Requirements

After setting up the environment, run this command in the virtual environment.

pip install -r setup/requirements.txt

---

Contributions: All of the code was written by Samiya Ali Zaidi.

Took help from this video to set up the environment.