CCMANet: A Cross-Layer Cascade Network With Multiattention Mechanisms for Remote Sensing Object Detection

Remote sensing object detection plays a vital role in both civilian applications and national defense security. However, remote sensing images typically exhibit characteristics such as wide coverage, significant variations in object scales, dense object distribution, and severe background interference. These factors greatly limit the applicability of existing detection methods in remote sensing scenarios. To address these challenges, this paper proposes a cross-layer cascade network with multi-attention mechanisms for remote sensing object detection (CCMANet). The proposed network incorporates different types of attention mechanisms at different stages to tackle background interference and dense multi-target detection, and leverages cross-layer cascading to progressively optimize feature representations, thereby achieving higher detection accuracy. Specifically, a multi-attention collaborative module is first introduced for feature filtering and suppression of complex backgrounds, highlighting useful remote sensing object features. Then, a maximum feature fusion module is employed in the feature fusion stage to enhance the diversity and representational capacity of the fused features. Finally, an improved dual-spatial pyramid pooling module combines two distinct spatial feature representations to further enrich target features in remote sensing images, ensuring that the dense and diverse remote sensing object information is preserved throughout the detection pipeline. Experiments on the DIOR, NWPU VHR-10, and RSOD datasets validate the effectiveness of the proposed method, achieving the highest mAP of 0.773, 0.952, and 0.973, respectively. Our code is available athttps://github.com/meijinlong/CCMANet.