Foreground Attention Loss and Attention-Guided Convolution for Remote Sensing Object Detection

Authors:
Publisher website
Google Scholar
Abstract
Remote sensing (RS) imagery contains vast backgrounds and densely packed, rotated small objects. One-stage rotated detectors therefore allocate most convolutional samples to background because they lack an explicit foreground discovery step. We close this gap with a lightweight discover→focus pipeline composed of two modules. First, a Foreground Attention Loss (FAL) supervises the classifier’s foreground heatmap by matching it to a rotated target–density map rendered from OBB annotations, giving the detector an explicit notion of “where the objects are.” Second, AGConv computes closed-form semantic–geometric dual offsets by fusing the learned foreground attention with anchor geometry, thereby concentrating sampling in salient regions while suppressing clutter without an extra offset-conv branch. This preserves one-stage efficiency while approximating the proposal-guided foreground extraction of two-stage detectors. Under identical ResNet–50+FPN settings, our single-stage detector attains 76.48% mAP on DOTA-v1.0, 68.65% on DOTA-v1.5, 67.2% on DIOR–R, and 97.28% (VOC12) / 90.82% (VOC07) on HRSC2016, with negligible parameter and latency overhead—demonstrating a strong accuracy–efficiency trade-off for complex RS scenes with dense, rotated targets.
Funding Information
  • National Natural Science Foundation of China (52274160)
  • National Natural Science Foundation of China (51874300)
  • Key Project of Innovation and Entrepreneurship Fund of Tiandi Science Technology Company Ltd. (2023-TD-ZD005-001)
  • Key Project of Innovation and Entrepreneurship Fund of Tiandi Science Technology Company Ltd. (2023-TD-ZD005-005)
  • Key Project of Innovation and Entrepreneurship Fund of Tiandi Science Technology Company Ltd. (2022-2-TD-ZD001)

This publication has 38 references indexed in Scilit: