AI Development Update - Fusheng Textile

Project Overview

The AI system developed in collaboration with Fusheng Textile Co., Ltd. aims to preserve traditional craftsmanship, streamline textile production, and enable designers to create high-quality textiles without in-depth technical knowledge. The system consists of three core components:

1. Textile Analysis and Property Extraction

   • A Convolutional Neural Network (CNN) analyzes textile images to extract key textile properties: structure, material, weight, and width.

2. ERP Production Optimization

   • A Gradient Boosting Model (GBM) or Neural Network will cross-reference extracted textile properties with Fusheng Textile’s ERP knitting production data (available yarns, machine capabilities, defect rates).

   • This allows the system to optimize resource allocation.

3. Jacquard Code Generation

   • An Encoder-Decoder Transformer Model will convert optimized textile properties into .RPP machine code for Mayer & Cie Circular Jacquard knitting machines, ensuring accurate production reducing production defects.

Value Proposition

• Preserve traditional craftsmanship by digitizing textile knitting process for future generations.

• One stop shop to empower designers to create physical textiles without extensive technical knowledge and connect with production facilities downstream.

• Reduce design turnaround time by automatically generating textile properties based on design.

• Optimize production efficiency through ERP-driven resource allocation.

• Minimize production defects by directly converting textile properties into machine code.

Current Roadmap To Develop the AI Model

1. Project Goal

• Input: Textile images

• Output: Predicted textile properties (structure, material, weight, width)

• Goal: Train an AI model that accurately predicts these properties from new textile images.

2. Data Preparation

Generating Segmentation Mask (Using OpenCV)

A segmentation mask acts as a guide to show what parts of the textile are important.

• Grayscale image conversion.

• Edge Detection to find patterns.

• Threshold to create binary mask.

Metadata Handling (Property Values):

• Convert categorical properties (structure, material) into one-hot encoded labels.

• Normalize numerical values (weight, width) using min-max scaling.

3. Model Architecture

Since the AI must handle both image-based classification and structured regression, the model will consist of:

• CNN Backbone (ResNet18) for extracting features from textile images.

• Dense layers for predicting structured textile properties (multi-task learning).

• Multi-output head for classification (structure, material) and regression (weight, width).

AI Model (Using PyTorch)

• Backbone CNN extracts features from textile images.

• Fully connected layers predict numerical and categorical textile properties.

• Multi-task learning handles classification & regression simultaneously.

Training Strategy:

• Training:

  • 10 epochs with validation monitoring to prevent overfitting.

  • Save best model based on validation loss.

5. Model Inference (Predicting New Textiles)

Once trained, the testing model will:

• Input a new textile image

• Predict textile properties (structure, material, weight, width)

---

Thanks to Fusheng Textile Co., Ltd. for supporting this ongoing research and development. For inquiries regarding the AI development please email me at awu@soooul.xyz. For any production inquiries feel free to contact Fusheng Textile in Taiwan at +886-3-212-2882