BRSTD: Bio-Inspired Remote Sensing Tiny Object Detection

作者：Sihan Huang, Chuan Lin, Xintong Jiang, Zhenshen Qu · 发表于：IEEE Transactions on Geoscience and Remote Sensing · 年份：2024 · DOI：10.1109/tgrs.2024.3470900 · 被引用次数：35 · 研究领域：Advanced Image and Video Retrieval Techniques、Smart Agriculture and AI

In aerial images captured by drones or satellite remote sensing images, object information is weak and difficult to distinguish from the background, with significant variations in object sizes. Physiological research indicates that the visual system can select visual stimuli through an attention mechanism, focusing resources on processing important information while suppressing less important information. Inspired by biological vision, this article designs an object detection network, named bio-inspired remote sensing tiny object detection (BRSTD). Drawing inspiration from the parallel pathways in biological vision and the antagonistic receptive field properties of X, Y, and W cells, we designed the XYW-Conv module with antagonistic receptive fields. This module enhances the contrast between tiny objects and their surrounding information, effectively extracting tiny object information from images. To further improve the backbone network’s ability to distinguish objects from the background, we designed the XYW-Attention and applied it to the designed Backbone. To better preserve tiny object information in the shallow feature layers and suppress large object and background information, we designed a feedback suppression attention module, top-down suppression attention (TDSA), at the connection between the neck and head parts, improving the model’s performance. Experiments show that with only 1.8 M parameters, BRSTD achieved 10.4 in terms of APvt and 23.6 in terms of average pre...