DSFB_Project

DSB4B - Data Labeling

This README describes the pipeline for annotating, labeling, and analyzing images and their annotated characters using a single notebook, final_notebook.ipynb. Each step in this process includes an explanation of the techniques used to achieve the desired outcomes.

---

Requirements

Before starting, ensure the following dependencies are installed:

• Python
• LabelImg (for bounding box annotation)
• A extension for running .ipynb files (simply download Anaconda or Miniconda)
• Additional Python packages (refer to final_notebook.ipynb for specific imports)

Organize your folder structure as follows:

• New_billets/New_billets/: Directory containing the original images to be annotated.
• New_billets/New_Billets_coords/: Directory where XML files containing bounding box annotations are saved (generated in LabelImg).
• annotations.csv: File storing image names and associated codes (created in Step 1).
• New_billets/New_billets_marked/: Directory where final labeled images will be saved.

---

Process Overview (Using final_notebook.ipynb)

Step 1: Data Annotation

1. Purpose: Load images and annotate each with a unique code.
2. Technique:
   • Manual Input and Data Storage: This step uses Python’s input() function in the notebook to manually enter a unique code for each image. The pandas library is used to structure and store this information in a DataFrame.
   • CSV Creation: The codes entered are then saved to annotations.csv using pandas.DataFrame.to_csv() to ensure a consistent format for later use.
3. Output:
   • An annotations.csv file containing each image filename and its associated code.

Step 2: Bounding Box Annotation with LabelImg

1. Purpose: Use LabelImg to create bounding boxes for each character in the code and label them in order.
2. Technique:
   • Manual Bounding Box Creation: LabelImg allows users to define regions of interest (bounding boxes) around each character. We save each character’s coordinates and label as an XML file.
   • Sequential Labeling: We label each character in ascending order to facilitate consistent annotation, which is crucial for the analysis in later steps.
3. Output:
   • XML files in the New_billets/New_Billets_coords/ directory, where each file contains bounding box coordinates and labels for the characters.

Step 3: Coordinate Addition and Labeling

1. Purpose: Combine data from the images, annotations.csv, and XML files to create labeled images.
2. Technique:
   • Data Parsing: Using xml.etree.ElementTree, the XML files are parsed to extract bounding box coordinates and labels.
   • Aligning Labels with Image Orientation:
     • To ensure labels align correctly with the orientation of each character, we calculate the center point of each bounding box.
     • We then compute the main orientation vector of the characters by analyzing the relative positioning of these center points.
     • Each character's center is projected onto this vector to establish a common axis for alignment.
     • Perpendicular vectors are calculated based on these projections, which serve as the baseline for displaying the labels, ensuring they are oriented consistently and are easy to read.
   • Image Processing with OpenCV: Using OpenCV, we add the center of the bounding box overlays and labels directly onto the images, highlighting each character’s position with visible annotations.
3. Output:
   • Final labeled images are saved in New_billets/New_billets_marked/, with the center of bounding boxes and character labels overlaid for easy visualization.

Step 4: Data Analysis in final_notebook.ipynb

1. Purpose: Analyze the annotated images by examining the frequency and positioning of labeled characters to gain insights into character distribution and positional patterns.

2. Techniques Used:

   • Data Aggregation and Parsing: Using pandas, data from annotations.csv is combined with positional data extracted from XML files. Each character’s position and label are stored in a DataFrame, which facilitates structured analysis.
   • Statistical Analysis:
     • Frequency counts and distribution patterns for each character are calculated using pandas and numpy. This allows for examining character occurrence trends, useful for understanding how characters are positioned across images.
   • Data Visualization: Visuals are generated with matplotlib and seaborn to display character frequency and positioning trends (e.g., histograms and scatter plots). These graphical representations make it easy to spot patterns in character locations and frequencies across the dataset.

3. Output:

   • Generated Files:
     • position_frequency.csv and char_frequency.csv: Summary files containing the frequency and position metrics for each character, which is useful for both statistical insights and quality control.

---

Running the Unified Notebook

1. Clone this repository or organize your files as specified in the folder structure.
2. Open final_notebook.ipynb and execute cells sequentially.
3. Review outputs and organize them for further analysis.

Example Output

• annotations.csv: Contains image names and respective codes.
• .xml files: Bounding box annotations from LabelImg for each character.
• Final labeled images: Images with annotated character coordinates in New_billets/New_billets_marked/.
• Data analysis outputs: Frequency statistics and charts.

---

This pipeline is designed to efficiently guide you through the entire process of image annotation, labeling and data analysis using a single notebook, with explanations provided for each technique applied.