AI Neural Painting
Summary
Associated with: University of São Paulo (USP).
Role: Student in Artificial Intelligence & Big Data (Postgraduate Diploma).
Where: São Paulo, Brazil.
When: 2023 - 2024.
Objective: Automate customer image processing for an e-commerce platform using computer vision and machine learning to reduce manual effort, time, and costs.
Contribution: Developed a full automation pipeline with OpenCV, Pillow, and Stylized Neural Painting, deployed on AWS, optimizing workflows from face detection to stylized effects.
Results:
• Reduced image processing time from ~26h to 15 minutes, achieving 100x faster performance.
• Lowered costs from $5 to $0.25 per image, making it 20x cheaper.
Project Description
This project introduces an automated system for preparing stylized images for e-commerce, specifically designed for a virtual store offering personalized gifts. The system leverages computer vision and machine learning techniques to optimize image processing, significantly reducing time and costs.
How It Works
The system automates image processing in several steps:
Image Upload: Customer uploads an image to be stylized and printed on a custom gift.
Face Detection: Automatically detects faces, re-aligns, and crops the image.
Background Removal: Removes background and enhances colors for visual quality.
Stylized Painting Effect: A neural network applies customizable painting effects, like oil painting.
Automation Server Pipeline: Processes each image in ~15 minutes at a cost of ~$0.25 per image.
Technical Information
The system incorporates several advanced techniques to enhance image processing. Neural Rendering employs an image-to-painting translation approach based on vector representations, creating stunning artistic transformations. Additionally, it utilizes Optimal Transport Loss, which leverages optimal transport theory to assess the similarity between the input image and the generated artwork. Finally, Differentiable Rendering calculates the derivative of the rendering process concerning the input parameters, enabling efficient optimization and fine-tuning of results.
Tools Used: OpenCV, Pillow, and Stylized Neural Painting (machine learning model).
Platform: Deployed on Amazon AWS for scalability and efficient batch processing.
Processing Time: Reduced from ~26 hours to 15 minutes per image.
Cost Efficiency: Lowers costs from $5 to $0.25 per image, significantly improving affordability.
Additional Information
The system not only automates the preparation process but also enhances scalability, enabling the store to handle higher volumes of orders without sacrificing quality. By eliminating manual intervention, the solution guarantees consistency, faster delivery, and reduced operational costs.