An Encoder-Decoder Architecture within a Classical Signal-Processing Framework for Real-Time Barcode Segmentation

Area of Science:

• Computer Vision
• Image Processing
• Augmented Reality

Background:

• Accurate barcode segmentation is essential for augmented reality (AR) applications.
• Existing methods may struggle with real-world image conditions like motion blur.

Purpose of the Study:

• To propose two novel methods for one-dimensional barcode segmentation.
• To achieve high accuracy and efficiency, particularly for AR applications.

Main Methods:

• Utilizing the partial discrete Radon transform as a core component.
• Developing a tile-based method for spatial and angle precision.
• Implementing an encoder-decoder network inspired by CNNs for segmentation without training.

Main Results:

• The encoder-decoder method achieves processing times faster than video acquisition on CPU for 1024x1024 images.
• Accuracy rivals state-of-the-art deep learning methods on standard datasets.
• The method excels with images exhibiting motion and lens blur, common in AR.

Conclusions:

• The proposed methods offer efficient and accurate barcode segmentation solutions.
• The encoder-decoder approach provides a competitive, non-training-based alternative for real-time AR.
• Implementations are provided for research and parallel processing on various CPUs.