Click-Pixel Cognition Fusion Network With Balanced Cut for Interactive Image Segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Interactive image segmentation (IIS) is crucial in fields like medicine and industry.
• Existing IIS methods struggle with core issues like pixel imbalance.
• Pixel imbalance stems from both pixel number and pixel difficulty.

Purpose of the Study:

• To propose a novel and unified network, Click-pixel Cognition Fusion with Balanced Cut (CCF-BC), to resolve pixel imbalance in IIS.
• To introduce a new loss function, Balanced Normalized Focal Loss (BNFL), that prioritizes hard-to-segment pixels.

Main Methods:

• Developed the Click-pixel Cognition Fusion (CCF) module, inspired by human cognition, to fuse click and visual information progressively.
• Proposed Balanced Normalized Focal Loss (BNFL), a general loss function that uses control coefficients to focus on positive and hard-to-segment pixels.
• Analyzed the theoretical relationship between BNFL, Focal loss, and BCE loss, showing they are special cases of BNFL.

Main Results:

• The CCF-BC method demonstrated superior performance compared to state-of-the-art methods on five benchmark datasets.
• The BNFL loss function effectively addresses pixel imbalance by focusing training on challenging samples.
• The CCF module successfully fuses click and visual information for improved segmentation.

Conclusions:

• The proposed CCF-BC network effectively solves the pixel imbalance problem in interactive image segmentation.
• The BNFL loss function offers a generalized approach to handling sample imbalance in segmentation tasks.
• The CCF-BC method represents a significant advancement in interactive image segmentation, with publicly available code for reproducibility.