Fourier Boundary Features Network With Wider Catchers for Glass Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Glass presents challenges in machine vision due to its transmittance and reflectance properties, blurring real-world boundaries.
• Accurate semantic segmentation of glass surfaces and penetration is crucial for reliable machine vision systems.
• Existing methods struggle to effectively segment glass boundaries and handle reflection noise.

Purpose of the Study:

• To propose a novel network, the Fourier Boundary Features Network with Wider Catchers (FBWC), for enhanced glass image segmentation.
• To address the challenges posed by glass transmittance and reflectance in machine vision tasks.
• To improve the accuracy and robustness of segmenting surfaces and penetrating glass in images.

Main Methods:

• Introduced the Fourier Boundary Features Network with Wider Catchers (FBWC), utilizing wide horizontal shallow branches for fine-grained boundary segmentation.
• Designed Wider Coarse-Catchers (WCC) for robust large-area segmentation and reduced feature over-extraction.
• Employed Cross Transpose Attention (CTA) to mitigate reflection noise and ensure complete boundary area segmentation.
• Developed a learnable Fourier Convolution Controller (FCC) for effective integration of multi-layer contextual information.

Main Results:

• The FBWC method demonstrated superior performance in glass image segmentation across three public datasets.
• Achieved better segmentation accuracy compared to existing state-of-the-art (SOTA) methods.
• Effectively cleared the boundary blurring caused by glass and reduced false positive feature captures.

Conclusions:

• The proposed FBWC network offers a significant advancement in glass image segmentation.
• The novel architectural components (WCC, CTA, FCC) effectively address challenges related to glass properties.
• FBWC provides a robust and accurate solution for machine vision tasks involving glass.