An Anti-background Interference Object Detection Method for Complex Remote Sensing Images

Object detection in remote sensing images differs from general object detection in that it faces numerous challenges such as false alarm interference. However, conventional object detection methods primarily focus on scale and orientation variations, insufficiently addressing issues such as object blurring and false alarms, which frequently occur in large-scale Earth observation scenarios. To address this issue, this paper proposes an anti-interference object-detection algorithm based on mask guidance and adaptive margin loss, termed MG2ADet. Specifically, we devise a mask guidance module that leverages pseudo-mask labels to guide network features towards focusing on foreground regions, thereby enhancing the discriminability of object features and suppressing background noise interference. Besides, to further differentiate object and background features, we designed an adaptive margin classification loss function that guides the detection model in learning strongly discriminative features. This classification loss adaptively adjusts the angular margin penalty based on the image quality score of samples, assigning varying weights to different difficult samples, and ultimately improving the detection model's ability to discriminate between blurred objects and similar distractors. Extensive experiments on public remote sensing datasets demonstrated the excellent detection performance of our algorithm in comparison with numerous existing detectors (both one-stage and two-stage detectors).