Improved UNet-Based Detection of 3D Cotton Cup Indentations and Analysis of Automatic Cutting Accuracy

Lin Liu, Xi‐Zhao Li, Hao Lv, Jianhuang Wang, Feipei Lai, Fangwei Zhao, Xibing LiProcesses 2025 · DOI

TL;DR

This paper presents an improved UNet-based deep learning model (UNet-IV) for automatically detecting indentations on 3D cotton cups, a crucial step for automated cutting. By integrating VGG16, Leaky-ReLU, ASPP, and an attention mechanism, the model achieves high accuracy in segmentation and significantly reduces cutting errors, paving the way for more efficient and cost-effective manufacturing of personalized textile products.

What Problem Does This Paper Address?

The production of personalized 3D cotton cups faces challenges due to the increasing variety of shapes and the limitations of traditional manual identification and cutting methods. Fixed molds are inefficient and costly, while manual processes are labor-intensive and prone to errors. Accurately identifying the precise contours of indentations on flexible materials like cotton cups is essential for automated cutting, but this is complicated by the material's deformability and complex shapes. Existing deep learning segmentation models, while powerful, may not be sufficiently robust for this specific, nuanced task, especially when aiming for high-precision industrial applications.

Methods

The core of this work is an enhanced UNet architecture, dubbed UNet-IV. Key improvements include:

  • VGG16 Backbone: Leveraging VGG16 as a pre-trained feature extractor for improved generalization and faster convergence.
  • Leaky-ReLU Activation: Replacing the standard ReLU activation function to mitigate the "dying ReLU" problem, ensuring better gradient flow and neuron activity, which is critical for learning fine indentation details.
  • Atrous Spatial Pyramid Pooling (ASPP): Incorporating ASPP to capture multi-scale contextual information by using dilated convolutions with varying rates, enhancing the model's ability to understand features at different scales.
  • ECA Attention Mechanism: Adding an Efficient Channel Attention module to allow the network to focus on more relevant feature channels, improving the detection of important indentation features and reducing the impact of background noise.

These components are integrated into the UNet's encoder-decoder structure to boost its segmentation accuracy and robustness for 3D cotton cup indentations.

Datasets

The study utilized a self-made dataset of 3D cotton cup images.

  • Image Collection: 632 images were collected using a Hikvision industrial camera under various lighting conditions, with different numbers of cups, and varying clarity levels to simulate real-world scenarios.
  • Dataset Split: From these, 500 images of well-rendered cups were selected for labeling. The dataset was split into training (70%), validation (20%), and testing (10%) sets.
  • Annotation: The "LabelMe" software was used for semantic segmentation annotation, requiring precise point-by-point labeling of the indentation contours for sub-pixel accuracy.

Key Insights and Ideas

  • Synergistic Improvement: The combination of VGG16, Leaky-ReLU, ASPP, and ECA modules proved to be more effective than individual enhancements, demonstrating a synergistic effect in improving segmentation performance.
  • Addressing "Dying ReLU": The adoption of Leaky-ReLU is crucial for learning the subtle, often negative-valued gradients associated with precise edge detection in flexible materials, preventing neuron inactivity.
  • Multi-Scale Context is Key: ASPP's ability to capture features at multiple scales is vital for accurately segmenting the varied shapes and sizes of cotton cup indentations, which can appear differently depending on the viewing angle and scale.
  • Attention for Focus: The ECA attention mechanism helps the model prioritize the critical indentation regions, effectively filtering out distracting background elements and improving the robustness of the segmentation.
  • Accuracy Matters for Manufacturing: Even seemingly small percentage improvements in segmentation metrics (like mIoU and mPA) translate into significant reductions in cutting errors and material waste in industrial settings, directly impacting cost and quality.

Key Results

  • Segmentation Performance: The proposed UNet-IV model achieved impressive results on the self-made dataset:
  • Precision: 99.53%
    • Recall: 99.69%
    • Mean Intersection over Union (mIoU): 99.18%
    • Mean Pixel Accuracy (mPA): 99.73%
  • Cutting Accuracy: After automated cutting using the segmented contours, the error analysis of 400 data points showed an average error of 0.20 mm ± 0.42 mm, meeting the stringent accuracy requirements for flexible material 3D cotton cups.
  • Comparison with Baselines: UNet-IV outperformed other established models like PSPNet, UNet++, and DeepLabV3+ in terms of mIoU and mPA, demonstrating its superior performance for this specific task.
  • Qualitative Improvement: Visual inspection of detection results showed that UNet-IV produced smoother, more complete edge contours compared to the baseline UNet, effectively avoiding artifacts like sawtooth edges and false detections.

Limitations and Future Work

  • Dataset Size and Diversity: While the dataset was designed to capture variability, future research could enrich it with samples under more extreme conditions (e.g., strong reflections, uneven textures) to further enhance model stability.
  • Model Lightweighting: For deployment on edge devices or high-speed production lines, efforts will be made to lightweight the UNet-IV model (e.g., using depth-wise separable convolutions or knowledge distillation).
  • Multi-Sensor Fusion: Exploring multi-sensor data fusion (e.g., incorporating 3D point clouds, force, and temperature data) could provide even richer information for more robust segmentation and control.
  • Broader Material Application: The approach could be extended to other multi-category flexible materials like sponge pads and non-woven fabrics.
  • Error Analysis Nuances: While the average error is low, the paper notes that error distribution varies slightly by cup type, with smaller cups showing a slightly larger margin of error. Further investigation into material properties and CNC control methodologies could refine this.

Summary updated

Abstract

With the advancement of intelligent technology and the rise in labor costs, manual identification and cutting of 3D cotton cup indentations can no longer meet modern demands. The increasing variety and shape of 3D cotton cups due to personalized requirements make the use of fixed molds for cutting inefficient, leading to a large number of molds and high costs. Therefore, this paper proposes a UNet-based indentation segmentation algorithm to automatically extract 3D cotton cup indentation data. By incorporating the VGG16 network and Leaky-ReLU activation function into the UNet model, the method improves the model’s generalization capability, convergence speed, detection speed, and reduces the risk of overfitting. Additionally, attention mechanisms and an Atrous Spatial Pyramid Pooling (ASPP) module are introduced to enhance feature extraction, improving the network’s spatial feature extraction ability. Experiments conducted on a self-made 3D cotton cup dataset demonstrate a precision of 99.53%, a recall of 99.69%, a mIoU of 99.18%, and an mPA of 99.73%, meeting practical application requirements. The extracted 3D cotton cup indentation contour data is automatically input into an intelligent CNC cutting machine to cut 3D cotton cup. The cutting results of 400 data points show an 0.20 mm ± 0.42 mm error, meeting the cutting accuracy requirements for flexible material 3D cotton cups. This study may serve as a reference for machine vision, image segmentation, improvements to deep learning architectures, and automated cutting machinery for flexible materials such as fabrics.
BibTeX
@inproceedings{liu2025improved,
  author = {Lin Liu and Xi‐Zhao Li and Hao Lv and Jianhuang Wang and Feipei Lai and Fangwei Zhao and Xibing Li and Xibing Li and Xibing Li},
  title = {{Improved UNet-Based Detection of 3D Cotton Cup Indentations and Analysis of Automatic Cutting Accuracy}},
  booktitle = {Processes},
  year = {2025},
  doi = {10.3390/pr13103144},
}
Figure