Localizing From Classification: Self-Directed Weakly Supervised Object Localization for Remote Sensing Images.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Object localization and detection in remote sensing images (RSIs) are crucial for various applications.
• Fully supervised methods demand extensive, costly instance-level annotations.
• Weakly supervised object localization (WSOL) offers a cost-effective alternative using image-level labels.

Purpose of the Study:

• To propose a novel self-directed weakly supervised strategy (SD-WSS) for object localization in RSIs.
• To enhance spatial feature extraction and address the discriminative region problem in WSOL.
• To reduce background interference and improve localization accuracy in complex RSIs.

Main Methods:

• Exploiting and enhancing the spatial feature extraction capabilities of classification models.
• Utilizing GradCAM++ to extract spatial location information for guiding the learning process.
• Designing a novel self-directed loss function to optimize model focus and mitigate background noise.

Main Results:

• The proposed SD-WSS method demonstrates superior performance compared to mainstream WSOL approaches.
• Extensive experiments were conducted on newly created WSOL benchmarks (C45V2 and PN2).
• The method effectively localizes objects of interest despite complex backgrounds in RSIs.

Conclusions:

• The SD-WSS strategy significantly advances weakly supervised object localization in remote sensing images.
• The approach offers a practical solution for reducing annotation efforts while maintaining high localization accuracy.
• The developed benchmarks and proposed method contribute to the advancement of WSOL research in the remote sensing domain.