An Improved Architecture for Semantic Segmentation Integrating Bidirectional Feature Pyramid Networks and Cross-Scale Feature Fusion for Effective Target Localization in High-Resolution Remote Sensing Imagery

Target detection in remote sensing imaging is impeded by difficulties in obtaining semantic information from diminutive objects and differentiating amongst like targets. This frequently results in inadequate performance from current detection techniques. This study presents IMPROVED-YOLO, an enhanced algorithm based on YOLOv9, specifically tailored for remote sensing imagery. IMPROVED-YOLO integrates a space-to-depth convolution module (SPD-Conv) to augment feature preservation, an attention mechanism in the backbone's final pooling layer for accurate target localization, a bidirectional feature pyramid network (Bi-FPN) for enhanced cross-scale feature fusion, and a decoupled head to optimize classification and regression tasks. Assessments of the VHR-10 and RSOD datasets reveal that IMPROVED-YOLO outperforms its predecessor, attaining mean Average Precision (mAP) scores of 93.5% and 93.9%, respectively, indicating a substantial enhancement over the baseline YOLOv9 algorithm.