Automatic Inside Point Localization with Deep Reinforcement Learning for Interactive Object Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Interactive image segmentation aims to precisely delineate objects in images.
• The Inside-Outside Guidance (IOG) algorithm shows strong performance but is sensitive to the initial inside point selection.
• Inconsistent inside point selection during training and testing degrades IOG performance.

Purpose of the Study:

• To develop a novel framework, the Inside Point Localization Network (IPL-Net), to automatically determine optimal inside points for IOG.
• To address the challenge of performance degradation caused by suboptimal inside point localization in interactive segmentation.

Main Methods:

• Proposed a deep reinforcement learning framework (IPL-Net) to infer the optimal inside point position.
• Formulated the inside point localization problem as a Markov Decision Process (MDP).
• Optimized the MDP using Dueling Double Deep Q-Network (D3QN) for robust localization.

Main Results:

• IPL-Net successfully localizes the inside point, enhancing the IOG algorithm's performance.
• The system automatically generates movement sequences to find the inside point after user input of two outside points.
• Trained and validated on the PASCAL dataset, demonstrating excellent performance.

Conclusions:

• IPL-Net effectively solves the inside point localization problem for interactive image segmentation.
• The deep reinforcement learning approach provides a viable solution for challenges in supervised learning data collection.
• The proposed method significantly improves the robustness and accuracy of the IOG interactive segmentation algorithm.