Convolutional Neural Network (CNN) for computer vision

CNN is a type of deep learning neural network that is very effective at computer vision. This repository will be a group of small projects regarding CNN for Computer vision using Keras and Tensorflow

Table of contents

• Keras Applications - Static Images
• MNIST classification digits
• MNIST classification Sign Language

---

1. Keras Applications - Static Images

Keras Applications are deep learning models that are made available alongside pre-trained weights. These models can be used for prediction, feature extraction, and fine-tuning - Source

Project Summary

Printed out top 5 prediction labels from models for 10 images. The results below states when the model predicted correctly out of the top 5.

Images used: (1) Pembroke Welsh Corgi, (2) Cocker Spaniel, (3) Giant Panda, (4) Hamster, (5) Hedgehog, (6) Brittany Spaniel, (7) Macaw, (8) Tabby cat, (9) Red-eyed Tree Frog, (10) French Bulldog

Model	Parameters	Known Top-5 Accuracy	Image Results
VGG19	143,667,240	0.900	7: top 1 2: top 2 1: no top 5
ResNet152	60,419,944	0.931	8: top 1 1: top 2 1: no top 5
Xception	22,910,480	0.945	8: top 1 1: top 4 1: not top 5
InceptionResNetV2	55,873,736	0.953	8: top 1 1: top 4 1: no top 5
NASNetLarge	88,949,818	0.960	8: top 1 1: top 4 1: no top 5

Analysis: Ranking Models

1. ResNet152
   
2. Xception, InceptionResNetV2, NASNetLarge
   
3. VGG19

---

2. Keras Built-in small Datasets: MNIST digits classification

The tf.keras.datasets module provide a few toy datasets (already-vectorized, in Numpy format) that can be used for debugging a model or creating simple code examples - Source

Project Summary

Creating a model based on the MNIST Dataset of grayscale image data with shapes and testing our model on test and real data.

Dependencies

Basic imports:

import numpy as np
import matplotlib.pyplot as plt
import os

ML imports:

import tensorflow as tf
from sklearn.preprocessing import MinMaxScaler
from tensorflow import keras
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Flatten
from tensorflow.keras.optimizers import SGD
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing import image
from tensorflow.keras.preprocessing.image import img_to_array
from IPython.display import Image, SVG

Steps

Load the Data

• Use tf.keras.datasets to load the 60,000 28x28 grayscale images of the 10 digits MNIST digits classification dataset
• This dataset has 2 splits: 'train' and 'test'

Explore datatset

There are 10,000 images in the test set
There are 60,000 images in the training set

Create Pipeline

• Normalize the training set and testing set
• Use MinMaxScaler to scale the numerical data
• Resize the training and testing from 28x28 to 784 pixels
• Shuffle and batch the data with batch_size = 32

Build and Train the Classifier

• Create a tf.keras Sequential model with the following layers:

  • Dense layer with 100 neurons and a relu activation function
  • Dropout layers with the dropout rate = 0.25 and 0.5 after each Dense layer. This prevents overfitting and produced the best results.

• Train the classifier

   model.compile(loss = 'categorical_crossentropy',
                 optimizer = SGD(0.01),
                 metrics = ['accuracy'])
  

• Fit the model

  • Use 10 Epochs
  • Used X_test, y_test as the validation data

   history = model.fit(X_train,
                       y_train,
                       batch_size = batch_size,
                       epochs = epochs,
                       verbose = 1,
   	                validation_data = (X_test, y_test))
  

• Test the model and print the loss and accuracy values

   Loss on the TEST Set: 0.11469
   Accuracy on the TEST Set: 0.9645
  

• Save the model

• Plot the loss and accuracy values achieved during training the the training and validation set

Analysis - Model

• Output prediction example of model(left) from image data(right)

• Output prediction example of model(left) from image data(right)

---

3. Kaggle dataset: MNIST classification Sign Language

Project Summary

Creating a model based on the Kaggle Dataset of an American Sign Language letter database of 24 hand gestures representing letters (excluding J and Z which require motion) and testing our model on test and real data.

Analysis

• Accuracy and Loss Chart

• Test Accuracy score: 0.7726

• Webcam Test

  • Top row (Sign language C, model correct)
  • Bottom row (Sign language A, model incorrect)

Limitations

• Model trained on sign language letters and not words or phrases
• Camera may not be getting correct angle that mirrors Sign Language MNIST dataset