A multiscale and cross-level feature fusion method for remote sensing image target detection

Abstract

Multiscale feature extraction and multilevel feature fusion are essential in the multiscale target detection of remote sensing images under complex backgrounds. However, mainstream feature extraction networks exhibit limited adaptability in their receptive fields and cannot accurately detect small targets with sparse features and large targets with blurred boundaries. Additionally, the feature fusion networks cannot effectively assess the importance of feature maps when integrating different depths of feature maps, resulting in an inaccurate representation of target area features. For this reason, we proposes a multiscale and cross-level feature fusion method for remote sensing image target detection. First, we construct a Multiscale Feature Extraction Network (MSFENet) to achieve comprehensive extraction of multiscale target features. We design a Scale-Adaptive Feature Extractor (SAFE) that encourages the model to realize multilevel synergistic cooperation within the sensing field, enabling accurate capture and efficient utilization of target feature information across different scales, thereby significantly enhancing the network’s capability to extract multiscale features. Additionally, in the deeper layers of the extraction network, we introduce a Cascading Context-Aware Module (CCAM) to enhance further the network’s ability to acquire deeper semantic richness while mitigating interference from background information. To fully exploit the latent information in feature maps at all levels of the backbone network, this study proposes a novel Cross-Level Feature Fusion Network (CLFFNet). This network incorporates a Sparse Feature Extraction Block (SFEB) to improve feature extraction quality. Through a carefully designed feature-adaptive cross-level interaction mechanism, the SFEB dynamically integrates the rich and critical detailed information with deep semantic information of the backbone network, significantly enhancing the integrity and discriminability of target feature information. The proposed module effectively improves the detection performance of small targets. The experimental results confirm the effectiveness and generalizability of the proposed method. Compared to existing target detection models, the proposed method achieves mAP values of 77.6 % and 90.1 % on the DIOR and HRRSD datasets, significantly outperforming the competitive models.