AI Postcard Detection & Warp Correction

Computer Vision • Perspective Correction • Warping Automation

Project Overview

This project delivers a full end-to end AI system designed to automatically detect postcards in raw photographs, correct geometric distortions, and apply neural warping to straighten curves and irregular edges, without ever opening Photoshop and without cropping or removing image content.
The final system reaches over 97% accuracy, meaning 97 out of 100 images require zero manual intervention, drastically reducing processing time from weeks to hours for large batches. The tool is now part of the client's internal production workflow, and it led to multiple follow-up projects due to performance and reliability.

Business Problem

The client needed to process extremely large batches of postcard photos often more than ten thousand at a time. Each image required manual perspective correction, edge straightening through warp adjustments, and visual cleanup inside Photoshop. Because the photos came from many different sources, they frequently contained blur, tilt, shadows, reflections and uneven lighting, making manual correction slow and exhausting.

A previous attempt using only OpenCV failed on most cases, delivering a 77 percent error rate. The client required a fully automated, reliable solution capable of handling real world distortions with professional level accuracy.

Solution

I developed a fully automated AI system capable of detecting the postcard inside each photo, correcting perspective distortion and applying a neural warping process to straighten curved or irregular edges without cropping any content. The model was trained on more than ten thousand manually annotated images and refined through multiple training cycles on A100 and H100 GPUs to achieve production level robustness.

The final tool runs as a standalone desktop application with a simple interface that processes entire folders at once and achieves more than 97% accuracy on new, unseen images provided by the client.

Project Milestones

Dataset Creation & Model Training

The first phase focused on building a high-quality dataset by manually annotating more than ten thousand postcard images. Once the dataset was balanced and cleaned, I trained multiple versions of YOLOv11 on H100 GPU instances. This stage also included generating precise masks and bounding shapes that would later guide the warping module.

Image Warping & Standalone Application

The second phase involved developing the neural warping engine capable of straightening curved or irregular edges while preserving one hundred percent of the postcard content. Perspective correction was fully automated, and the warping logic was refined to handle even difficult lighting and geometric variations. I then integrated the entire pipeline into a standalone desktop application. A rejection mechanism was included to isolate ambiguous or multi-postcard inputs, and the final delivery was a production-ready tool used directly within the client’s workflow.

Project Conclusion

The system consistently produced clean, rectangular outputs in more than 97% of cases, meaning almost all images required no manual correction. The reject rate remained below three percent, meeting the client’s benchmark, and the total processing time for large batches dropped from several weeks to just a few minutes. Because the entire workflow is fully automated and does not involve Photoshop at any stage, the client was able to adopt the tool in production immediately and later commissioned additional AI projects based on its performance.

Impact

Eliminated manual Photoshop workload
Reduced operational time by 99%
Provided a scalable AI solution used across multiple batches
Enabled the client to process postcard inventories at industrial scale

⭐ Client Reviews

Page updated

Google Sites

Report abuse