Week 4 Record

 

  Week 4: Gravestone OCR & Final Pipeline Integration


21/02/2025


Summary of Activities

1.    Applying the Pipeline to Gravestones

•  Received a set of gravestone images and used our modular functions to organize (IR vs. visual), enhance, segment, and then run OCR.

2.    OCR Challenges

•  Gravestones often have low contrast and weathered text, causing Tesseract/Easy OCR to misinterpret partial letters, moss, cracks, and unusual fonts.

•  Tried additional morphological operations and stricter threshold parameters, but results still varied by image condition.

3.    Refined Code for OCR

•  Implemented a final function, RefinedEdgeOverlayAndOCR, which merged thresholding + edge detection and thickened letter boundaries in red.

•  This approach slightly improved OCR performance but did not fully solve recognition errors on heavily weathered stones.




Key Issues Encountered

•  High Noise on Gravestones: Weathering, moss, and stone texture introduced significant noise that standard filtering struggled to remove.

•  Letters & Background in Same Color Range: The engraved text often had the same or very similar tonal range as the stone, making segmentation highly challenging for OCR.

•  Partial or Merged Letters: Erosion created irregular shapes that standard OCR engines could not correctly interpret.

 

Final Observations

Unfortunately, the final results were not as satisfying as we had hoped. The gravestones’ high noise levels, combined with text having nearly the same color as the background, severely impacted OCR accuracy. Despite improved boundary detection and morphological filtering, many letters remained indistinguishable from the stone surface.

 

Proposed Solutions

•  Short-Term: Further refine preprocessing (e.g. super-resolution, advanced morpho- logical filters). Restrict OCR to known character sets or specific page segmentation modes.

•  Long-Term: Train a custom deep learning OCR model specifically for engraved text (e.g. CRNN, Vision Transformers). Investigate specialized imaging methods (e.g. more IR captures) that might reveal hidden or faint text more clearly.

 

Progress Assessment

1.    We successfully built a modular MATLAB pipeline:

•  OrganizeImages.m for categorizing images by IR vs. Visual vs. Identifiable.

•  EnhanceAndRestore.m for median filtering, histogram equalization, and sharpening.

•  FeatureExtractionAndSegmentation.m for basic Canny edges and adaptive thresholding.

•  RefinedEdgeOverlayAndOCR.m to merge threshold & edge data and attempt OCR.

•  MainPipeline.m to tie it all together and automate the process.

2.    Demonstrated the pipeline on IR images (GasTube detection) and on visual gravestone images (OCR focus).

3.    Learned the limitations of off-the-shelf OCR engines on weathered or low-contrast text.

 

Concluding Thoughts & Future Work

Over four weeks, we progressed from initial planning to a functioning MATLAB-based im- age processing/recognition pipeline. While we had reasonable success with clearer subjects (like gas tubes in IR images), weathered gravestones proved significantly more challenging, especially when stone and lettering shared similar color intensities.

 

What’s Next?

•  Advanced OCR Techniques: Custom-trained models on a representative dataset of engraved text.

•  Deeper Preprocessing: Try advanced de-shadowing, super-resolution, or specialized morphological operations to separate text from background.

•  Larger, Diverse Dataset: Gathering more gravestone images with varied lighting, angles, and text styles could help improve OCR model robustness.







Comments

Post a Comment

Popular posts from this blog

Week 1 Record

  Week 1: Project Kick-Off & Planning 31/01/2025 Summary of Activities 1.     Initial Project Scope •   Held a meeting to clarify overall project goals: process both IR and visual images, enhance them, and extract meaningful features or text (especially for weathered gravestones). •   Discussed potential segmentation methods, OCR technologies, and the importance of robust pre-processing steps. 2.     Roles Acquiring and organizing images (IR vs.   Visual, Identifiable vs. Non-identifiable).        Image enhancement and feature extraction.      Recognition model development and testing. 3.     Project Aims •   Created a 4-week timeline with clear milestones: –     Week 1: Project setup, role assignments, initial code planning. –     Week 2: Acquire images...
Read more

Week 2 Record

      Week 2: Delays in Image Acquisition & Initial Tests 07/02/2025 Summary of Activities 1.     Image Delay •   Repeatedly reached out for the official IR and visual image bank but still had no official delivery. 2.     Preprocessing Brainstorming •   Despite limited real data, we discussed potential noise reduction, contrast adjustment, and morphological operations in MATLAB. 3.     Test Cases •   Downloaded or generated placeholder/sample images to test our preliminary code structure, ensuring each function worked smoothly.   Challenges/Concerns •   Scheduling Conflicts: Some team members had overlapping lab sessions. We used a shared calendar to align team coding sessions. •   Resource Availability: Confirmed MATLAB license but uncertain about OCR tool- boxes (e.g., Tesseract or Easy OCR integration).
Read more

Week 3 Record

Image
                   Week 3: IR Image Processing & Preliminary MAT- LAB Code 14/02/2025 Summary of Activities 1.     IR Image Analysis •   Finally received a few infrared images (e.g., “IR 00345.jpg”) and tested our Gas- Tube boundary detection script. •   Implemented Gaussian smoothing, contrast enhancement, and Otsu thresholding to isolate key regions (the gas tube). 2.     Segmentation Approaches •   Explored both threshold-based segmentation and Canny edge detection. •   Found that morphological operations (imclose, imopen, imfill) were critical to suppress noise and highlight main object boundaries. 3.     Code Organization •   Consolidated our functions into single MATLAB scripts or Live Scripts. Added thorough inline comments and verified toolbox requirements. Challenges/Concerns •   MATLAB T...
Read more