Remote Sensing Image Object Detection Based on Adaptive Feature Weighting

The broad field of view inherent to remote sensing images offers significant advantages for large-scale monitoring tasks， enabling comprehensive surveillance over vast geographical areas with a single acquisition. However， this advantage is often offset by the presence of complex backgrounds， such as cloud cover， shadows， and varying terrain textures， which frequently introduce channel-level noise that interferes with target features. This interference significantly compromises the robustness and reliability of detection models， particularly in adverse weather conditions or densely cluttered environments. Additionally， the substantial variation in target sizes， ranging from small objects like vehicles and helicopters to large structures such as ports and bridges， poses a significant challenge. Detection models struggle to simultaneously capture high-resolution details necessary for identifying small targets and low-resolution semantic information essential for recognizing larger objects， leading to unbalanced detection performance across different scales.To address the aforementioned challenges， this paper proposes a remote sensing image object detection algorithm based on adaptive feature weighting， built upon the Oriented R-CNN framework. The method introduces two core modules： the Adaptive Channel Weighting Module （ACWM） and the Adaptive Multi-Scale Fusion Pyramid （AMFP）， aimed at tackling background interference and multi-scale target detection issues. The ACWM module is applied after the STAGE 2 and STAGE 3 phases of the feature extraction network to adaptively adjust channel weights， a critical step for enhancing feature representation. By analyzing statistical correlations between channels using the Correlation-Aware Covariance Weighting （CACW） method， it generates dynamic channel weights to enhance target-related features while suppressing background noise and redundant information， balancing the expression of different channels and thereby improving the feature extraction network's ability to extract features from rotated targets. Subsequently， the AMFP module processes the multi-scale features （C2， C3， C4， and C5） extracted by the feature extraction network， employing a hierarchical approach for feature integration. The Multi-scale Perception and Context Integration Module （MPCIM） fuses global semantic features from deep layers with edge and positional details from shallow layers， effectively addressing the issue of small targets being obscured. The Adaptive Layer Weighting Module （ALWM） optimizes the fusion quality of multi-scale features through adaptive layer-wise weighting， enhancing the model's capability to handle diverse target scales ranging from small vehicles to large ports by dynamically adjusting the contribution of each feature level， thus improving adaptability to various target scales. The optimized features （P2， P3， P4， and P5） are then fed into the Oriented RPN and Oriented RCNN head， which， based on the Oriented R-CNN framework， generates high-precision rotated bounding boxes to complete classification and localization tasks.To validate the effectiveness of the proposed algorithm， extensive experiments were conducted on the DOTA1.0 and DIOR-R datasets. Ablation studies were performed to assess the contributions of the ACWM and AMFP modules. Results indicate that the ACWM module alone improves the mean Average Precision （mAP） by 0.97% on DOTA1.0 and 1.73% on DIOR-R， while the AMFP module contributes an additional 1.30% and 2.01% improvement， respectively. When both modules are combined， the mAP reaches 78.23% on DOTA1.0 and 68.17% on DIOR-R， representing overall enhancements of 2.36% and 3.87% compared to the baseline Oriented R-CNN model. Visualization comparisons further demonstrate the superiority of the proposed method， showing a significant reduction in missed detections and false positives for small targets， particularly in categories such as Small Vehicle， Swimming Pool， Helicopter， and Storage Tank. The method also exhibits improved localization accuracy and better contour fitting for rotated objects， especially in complex backgrounds with varying angles and dense distributions. The experimental results highlight the robustness and adaptability of the proposed algorithm across diverse remote sensing scenarios.Compared to state-of-the-art methods such as Faster R-CNN， RoI Transformer， and ARC， the proposed approach achieves superior performance in handling small targets and complex orientations， as evidenced by its higher mAP and enhanced visual detection outcomes. However， limitations remain， including potential performance degradation in extremely cluttered scenes or with ultra-small targets， suggesting avenues for future optimization.