ODOMETER CLASSIFICATION USING TRODO DATASET

Project Overview

The "trodo-odometer-classification" project aims to classify odometer types and extract mileage information using the TRODO dataset. The project utilizes various machine learning techniques, including K-Nearest Neighbors (KNN), Random Forest Classifier, and Convolutional Neural Networks (CNN) to accurately classify odometer images as either analog or digital.

The entire workflow is encapsulated in the Jupyter Notebook ./odometer_classification_trodo.ipynb, which provides step-by-step instructions, code, and outputs for replicating the study.

Usage

To replicate the project, follow these steps:

1. Download and Extract the Dataset

You can access the TRODO dataset from the following link: TRODO Dataset

Mount the Google Drive to access the dataset:

from google.colab import drive
drive.mount('/content/drive')

2. Explore the Dataset

Import necessary libraries and set the dataset folder path:

import os
import json
import cv2

dataset_folder_path = '/content/drive/MyDrive/trodo-v01'
groundtruth_file_path = os.path.join(dataset_folder_path, 'ground truth', 'groundtruth.json')
annotations_file_path = os.path.join(dataset_folder_path, 'pascal voc 1.1', 'Annotations')
images_file_path = os.path.join(dataset_folder_path, 'images')

Open and inspect the groundtruth.json file:

with open(groundtruth_file_path, 'r') as f:
    groundtruth_data = json.load(f)

groundtruth_data['odometers'][0].keys()

3. Data Preprocessing

Install the tqdm library for tracking the preprocessing progress.

Preprocess the images by extracting and resizing the odometer part:

import os
import cv2
import xml.etree.ElementTree as ET

Convert string labels to numeric labels using LabelEncoder:

from sklearn.preprocessing import LabelEncoder

label_encoder = LabelEncoder()

Save and load the preprocessed data to avoid re-running preprocessing every time:

import pickle

Split the dataset into training and testing sets:

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

4. Train and Evaluate the Models

KNeighborsClassifier

Train the K-Nearest Neighbors Classifier and find the best parameters using GridSearch:

from sklearn.neighbors import KNeighborsClassifier
from sklearn.model_selection import GridSearchCV

param_grid = [{'weights': ["uniform", "distance"], 'n_neighbors': [3, 4, 5, 6]}]

Evaluate the model:

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score

Random Forest Classifier

Train the Random Forest Classifier:

from sklearn.ensemble import RandomForestClassifier

Evaluate the model:

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, confusion_matrix

CNN

Train a Convolutional Neural Network with Early Stopping:

from tensorflow.keras.models import load_model

Load the best model weights and summarize the model:

model.load_weights('best_model.h5')
model.summary()

Dependencies

• Python
• Jupyter Notebook
• Google Colab (for mounting drive)
• OpenCV
• NumPy
• scikit-learn
• TensorFlow
• tqdm
• matplotlib
• seaborn
• pickle

Ensure you have all required libraries installed to run the notebook successfully.