改進RRPN模型的遙感圖像目標檢測

摘" 要： 針對遙感目標背景復雜、易受外界環境干擾，傳統方法無法滿足復雜場景下的檢測高精度與實時性要求的問題，提出基于改進RRPN模型的遙感圖像目標檢測方法。首先，將特征金字塔（FPN）架構引入到了模型的殘差網絡中，使得遙感圖像的高、低層特征得到了有效融合；其次，在特征提取網絡中添加了通道和空間相融合的注意力機制（CBAM），提升了模型在遙感圖像目標特征提取方面的跨通道和空間處理能力；此外，將剔除重疊建議框時的原始NMS算法優化為DIoU?NMS算法，綜合考慮遙感圖像候選框之間的重疊度、距離、尺度大小等因素，使目標框的回歸過程更加穩定。對比實驗與消融實驗顯示，所提方法在公共數據集DOTA和HRSC2016上獲得的平均精度均值mAP分別可高達77.30%、90.24%，較原始RRPN模型分別提高了8.29%、11.16%，且優于其他幾種較新的經典模型，表明所提方法對于復雜環境下的遙感圖像目標檢測是合理且有效的。

關鍵詞： 目標檢測； 遙感圖像； 帶旋轉的候選框算法； 卷積通道注意力模塊； DIoU?NMS； 特征金字塔； DOTA；HRSC2016數據集

中圖分類號： TN911.73?34； TP391.4" " " " " " " " "文獻標識碼： A" " " " " " " " " 文章編號： 1004?373X（2025）01?0008?09

Remote sensing image object detection based on improved RRPN model

LU Xiaobo， GUO Yanguang， XIN Chunhua

（Department of Computer Technology and Information Management， Inner Mongolia Agricultural University， Baotou 010010， China）

Abstract： The background of the remote sensing object is complex. In addition， the remote sensing object is susceptible to external environment， so the traditional methods fail to meet the requirements of high precision and real?time detection in complex scenes. In view of this， the paper proposes a remote sensing image object detection method based on the improved RRPN model. The framework of feature pyramid network （FPN） is introduced into the residual network of the model， which enabled the effective fusion of high? and low?level features of remote sensing images. The convolutional block attention mechanism （CBAM） combining channel and space is incorporated into the feature extraction network， so as to improve the cross?channel and spatial processing capability of the model in the feature extraction of remote sensing image object. In addition， the original NMS （non?maximum suppression） algorithm is optimized into DIoU?NMS algorithm for eliminating overlapping object frames， and the overlap， distance， scale and other factors among the candidate frames of remote sensing images are taken into account comprehensively， so as to make the regression of object frames more stable. In the comparative and ablation experiments， it is shown that the proposed method achieves mAP （mean average precision） of 77.30% and 90.24% on the public datasets DOTA and HRSC2016， respectively， which are 8.29% and 11.16% higher than that of the original RRPN （rotation region proposal network） model， and it is better than that of the other advanced classical models. This indicates that the proposed method is reasonable and effective for the object detection of remote sensing images in complex environments.

Keywords： object detection; remote sensing image; RRPN; CBAM; DIoU?NMS; FPN; DOTA; HRSC2016 dataset

0" 引" 言

近年來，隨著遙感技術在諸多領域的蓬勃發展，高分辨率光學遙感圖像憑借其分布面積大、范圍廣等特點，受到國內外研究專家的廣泛關注[1]。遙感圖像目標檢測[2?3]作為其中的關鍵性任務，檢測性能的優劣對于船舶等的動態監測、交通秩序的監測與管理、城市建筑物的規劃等工作造成了直接影響。此外，由于遙感圖像自身具有目標方向不確定、排列密集、尺度不一等特點，這為復雜場景下的遙感圖像檢測工作帶來了挑戰。……