EnNet: Enhanced Interactive Information Network with Zero-Order Optimization.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Interactive image segmentation is crucial for generating high-quality datasets for deep learning.
• Current methods struggle with insufficient interaction information and high optimization costs, leading to suboptimal results and increased computational burden.

Purpose of the Study:

• To address the limitations of existing interactive image segmentation methods by improving the mining of interaction information within network architecture and optimization.
• To enhance the representativeness of interactive regions and mitigate the weakening of interaction information through network hierarchy.

Main Methods:

• Proposes EnNet, a novel network architecture utilizing attention mechanisms to integrate user interaction information across the entire image.
• Incorporates interaction information twice in a coarse-to-fine design to strengthen its impact.
• Introduces a zero-order optimization method for the initial training iterations to reduce computational overhead with minimal accuracy loss.

Main Results:

• EnNet effectively integrates user interaction information, improving feature representation in interactive regions.
• The zero-order optimization approach significantly reduces computational costs during early training stages.
• Experimental validation on GrabCut, Berkeley, DAVIS, and SBD datasets demonstrates superior performance, surpassing RITM by 0.35 in average NOC@90.

Conclusions:

• EnNet offers an effective solution for interactive image segmentation, balancing accuracy and computational efficiency.
• The proposed methods enhance the utilization of interaction information and optimize the training process.
• This work contributes to accelerating the generation of annotated datasets for deep learning applications.