Explicit–Implicit Collaborative Attention Guided Weakly Supervised Object Detection in Remote Sensing Images

Abstract

The task of detecting objects in remote sensing imagery is particularly challenging due to dramatic scale variability, cluttered backgrounds, and densely packed instances. To circumvent the prohibitive cost of instance-level annotation, weakly supervised object detection (WSOD) techniques that rely solely on image-level labels have attracted significant research interest. However, existing WSOD methods are prone to the part domination problem, i.e., the model overfits to the most discriminative local regions rather than the complete objects, resulting in detection accuracy significantly lower than that of fully supervised methods. To address the part domination problem, this article proposes a WSOD method for remote sensing images based on explicit–implicit collaborative attention guidance, which synergistically alleviates this problem from two levels: explicit structural design and implicit constraint learning. At the explicit level, a bottom-up aggregation attention enhancement module is designed, which integrates multilevel feature concatenation with a frequency-domain Transformer to fuse shallow spatial details and deep semantic features, thereby enhancing the positional awareness of complete objects in deep features. At the implicit level, a comprehensive attention guidance mechanism is introduced, which constructs a comprehensive attention target using attention maps from different network layers and imposes a consistency loss to force the attention maps from each perspective to approach this comprehensive target, thus suppressing the positive feedback loop of incomplete pseudolabels. Experimental results on two public remote sensing datasets, DIOR and NWPU VHR-10.v2, show that the proposed method achieves mean average precision of 30.82% and 72.24%, respectively, outperforming existing mainstream WSOD methods and effectively alleviating the part domination problem.