XCotton: Advancing AI-Enabled Hardware/Software Integrated System for Foreign Fiber Cleaning

Gaowei Zhang, Wei Wang, Yusheng Peng, Yi WangAAAI 2025 · DOI

TL;DR

This paper introduces XCotton, an AI-powered system that integrates advanced hardware and software to automatically detect and remove foreign fibers from cotton. This innovation significantly boosts cleaning efficiency and reduces energy consumption, offering substantial economic and social benefits, particularly for developing countries.

What Problem Does This Paper Address?

Cotton is a vital global commodity, but its processing is plagued by contamination from foreign fibers like plastic, hair, and synthetic materials. These contaminants are difficult to remove, fragment into smaller pieces during processing, and drastically reduce the quality and value of cotton products. Traditional cleaning methods are labor-intensive, prone to human error, and increasingly insufficient as mechanized harvesting introduces more diverse contaminants. This poses a significant challenge for developing countries heavily reliant on cotton exports, hindering their ability to move up the value chain. The paper addresses the need for an automated, efficient, and accurate solution to this persistent problem.

Methods

The XCotton system is a hardware-software co-design featuring a three-stage approach:

  • Pre-Cleaning (Stage One): This stage uses mechanical methods like vibrating screens, air jets, and gravity separators to remove larger impurities such as soil and debris, preparing the cotton for finer cleaning.
  • Foreign Fiber Detection (Stage Two): This is the core of the system. It employs a line-scan camera paired with three distinct light sources: white light for colored fibers, and red and green lasers to enhance contrast for low-visibility contaminants like white plastic films. The system uses a novel few-shot foreign fiber detection (FSFFD) network, which adapts the CLIP model. This network utilizes "object-aware text prompts" to learn generalized embeddings for normal and abnormal (foreign fiber) classes, reducing the need for extensive manual prompt engineering. It incorporates a joint optimization strategy that balances global and local context losses to improve detection accuracy.
  • Packaging (Stage Three): This stage, developed by their collaborator, uses air jets to remove detected foreign fibers and package the cleaned cotton.

A key technical aspect is the field correction and calibration algorithm for the line-scan camera, ensuring consistent image quality by correcting for lens distortion and misalignment.

Datasets

The authors created a dataset by continuously capturing over 10,000 images from actual production lines. Approximately 3,000 of these images contain foreign fibers. Each image has a resolution of 4096 × 200 pixels. A subset of 100 images containing foreign fibers was made publicly available.

Key Insights and Ideas

  • Synergy of Hardware and AI: The paper highlights the power of integrating specialized hardware (line-scan camera, specific lighting) with advanced AI (few-shot learning, object-aware prompts) for a highly effective solution.
  • Few-Shot Learning for Niche Problems: Adapting CLIP for few-shot foreign fiber detection demonstrates its potential beyond natural language and image tasks, proving effective even with limited specific training data for this industrial application.
  • Object-Aware Prompts: The novel text prompt templates that incorporate prior semantic knowledge significantly improve the model's ability to discriminate between cotton and foreign fibers, reducing reliance on extensive manual feature engineering.
  • Global + Local Context Optimization: Combining global and local context losses in the AI model is crucial for achieving superior accuracy, capturing both broad anomaly semantics and fine-grained local features.
  • Economic and Social Impact: The system's success isn't just technical; it has a tangible impact on the cotton industry, especially in developing nations, by increasing efficiency, reducing costs, and enabling industry upgrades.

Key Results

  • Efficiency: XCotton achieves a cleaning efficiency of 1000 kg/h, a 43% improvement over previous generations.
  • Energy Consumption: It uses only 14 kWh of energy, a 63% reduction.
  • Accuracy: The system achieves an accuracy of 92.9% and an IoU (Intersection over Union) of 86.2%, significantly outperforming traditional methods like SVM, UNet, and even a standard CLIP approach.
  • Commercial Success: ZhongMian has received over 100 orders for the new machine, with an estimated annual revenue of 25.2 million CNY (approximately 3.5 million USD). Over 40% of these orders are from African and Central Asian clients.
  • Cost Savings: For clients, one XCotton machine can reduce manual labor by about 70 workers, saving approximately 0.5 million CNY annually in labor costs, often exceeding the machine's purchase price.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the system is highly effective, the computational resources required for training and deployment can be significant, especially for real-time applications. The effectiveness of machine learning models is also heavily dependent on the quality and quantity of labeled training data.
  • Future Work: The authors' work focuses on advancing AI-enabled hardware/software integration for foreign fiber cleaning. Future directions could involve further optimizing the computational efficiency, exploring even more diverse types of contaminants, and expanding the system's deployment to a wider range of textile processing applications. The success in LDCs also suggests potential for further research into tailoring such technologies for specific regional needs and economic contexts.

Summary updated

BibTeX
@inproceedings{zhang2025xcotton,
  author = {Gaowei Zhang and Wei Wang and Yusheng Peng and Yi Wang},
  title = {{XCotton: Advancing AI-Enabled Hardware/Software Integrated System for Foreign Fiber Cleaning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35146},
}
Figure