Accurate stacked-sheet counting method based on deep learning.

Area of Science:

• Industrial engineering
• Computer vision
• Deep learning

Background:

• Accurate counting of laminated sheets (e.g., packing, printing) is crucial for industrial economic costs.
• Traditional image processing methods struggle with varying sheet thicknesses, adhesion, breakage, and low contrast.

Purpose of the Study:

• To propose a novel stacked-sheet counting method utilizing a deep learning approach.
• To address the limitations of traditional image processing in counting laminated sheets.

Main Methods:

• A U-Net deep learning architecture was employed for semantic segmentation of sheet centerlines.
• A custom dataset of stack side images was collected and utilized.
• The model was trained on small image patches and tested on large images for segmentation.
• Pixel classification using multi-layer convolution and deconvolution identified sheet centerlines.

Main Results:

• The deep learning model successfully segmented centerlines of sheets from stack side images.
• The number of sheets was accurately determined by calculating the median value of centerline points.
• The proposed method demonstrated higher accuracy and a lower error rate compared to traditional algorithms in real-world experiments.

Conclusions:

• The U-Net based deep learning approach offers a robust solution for laminated sheet counting.
• This method overcomes challenges posed by sheet variations and low contrast, improving counting accuracy and efficiency.