Image Classification using CNN TensorFlow OpenCV

CNN Image Classification using TensorFlow

This guide will help you navigate and learn CNN Image Classification using TensorFlow. This is an Introductory Codebase. For deeper knowledge visite GitHub : DeepLearning-ImageClassification-Toolkit

TensorFlow is Required to run this Code.

Install using THIS Guide. #### Make sure to inastall necessary dependencies by running this command : pip install -r requirements.txt

Understanding the Functionalities

1. PreProcessing

Load Dataset from ‘DATA_DIR’ Directory
Creates DataFrame containing Full Paths of Images and their Class Labels
(Change as per Requirement) Rescale Images to Computationally Efficient Resolution
(Optional but Recommended) Extracts Largest Object from Image using ‘image_processing’ Function
- Leverages Parallel Processing for Faster Results
Compares Original and Rescaled+Processed Image SIde-By-Side to make necessary changes
Converts Processed Images to NumPy Array and Exports as Pickle File
- Verifies If Exported Pickle File is Appropriate through 10 Random Samples
(Optional) Merge Certain Class Lables Together
Split Data for Training, Testing, Validation with Stratify to ensure data balancing
- Verify if Split is Appropriate through 2 random samples
(Optional) Perform Random Oversampling on Data to reduce Biasness
- Verify if Oversampling is Appropriate through 2 random samples
Perform One-Hot-Encoding of Class Labels
Training, Testing, Validation Data and One-Hot-Encoding is Exported as Pickle Files

2. Training CNN ResNet50

Training, Testing, Validation Data and One-Hot-Encoding are Imported
- All Data is converted to TensorFlow Format
Learning Rate Scheduler is Defined (Change If Desired)
Stochastic Gradient Descent with Momentum is Used as Optimizer (Change If Desired)
Added Data Augmentation Techniques to improve Model Learning (Change If Desired)
Base Model (CNN ResNet50) is Loded from TensorFlow Library
- Custom Optimal Changes have been made to the Structure
- Final Model is Compiled
Final Model is Trained
- Final Model with Lowest Validation Loss is Exported as a ‘.h5’ file
Traning Time (In Seconds) is Displayed

3. Verification and Confusion Matrix

Trained Model and One-Hot-Encoding are Imported
Entire Dataset is Run through the Trained Model to get Ground Truth of Accuracy
(Optional) Incorrectly Classified Image Files will be copied to a seprate folder with detected class label
Ground Truth Classification Confusion Matrix is Created
- (Optional) Confusion Matrix can be Exported as a ‘.png’ file

HelperFunctions

Function to display 2 images side-by-side on screen
Function to Extract largest object from souruce image

Source Code: HelperFunctions.py'''
Source : https://github.com/iSiddharth20/DeepLearning-ImageClassification-Toolkit
'''

import matplotlib.pyplot as plt
import cv2
import numpy as np

'''
Helper Function 
    - Used to Show 2 Images Side-By-Side
'''
def images_on_side(img_1,label_1,img_2,label_2):
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 24))
    ax1.imshow(img_1)
    ax1.set_title(label_1)
    ax2.imshow(img_2)
    ax2.set_title(label_2)
    plt.show()

'''
Helper Function 
    - Used to Extract Object from Image
'''
def image_processing(image_path):
    # Read the image
    img = cv2.imread(image_path)
    # Convert image to grayscale
    gray_img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    # Threshold the image to get a binary mask
    _, thresholded = cv2.threshold(gray_img, 135, 255, cv2.THRESH_BINARY)
    # Perform morphological closing
    kernel_size = 5
    kernel = np.ones((kernel_size, kernel_size), np.uint8)
    closed_img = cv2.morphologyEx(thresholded, cv2.MORPH_CLOSE, kernel)
    # Find contours
    contours, _ = cv2.findContours(closed_img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    # Identify the largest contour
    largest_contour = max(contours, key=cv2.contourArea)
    # Create an empty mask and draw the largest contour onto it
    contour_mask = np.zeros_like(thresholded)
    cv2.drawContours(contour_mask, [largest_contour], -1, (255), thickness=cv2.FILLED)
    # Dilate the mask slightly
    dilated_mask = cv2.dilate(contour_mask, kernel, iterations=1)
    # Use the mask to extract the largest object from the original image
    extracted_object = cv2.bitwise_and(img, img, mask=dilated_mask)
    return extracted_object