LCMCA-Net: A Large-Kernel Contour Attention and Multiscale Corner Alignment Network for Remote Sensing Object Detection
Abstract
Remote sensing object detection plays a crucial role in national defense and security applications. Although anchor-free methods such as CenterNet++ achieve competitive recall and low false-positive rates, they remain limited by restricted global receptive fields and insufficient exploitation of geometric contour information. In addition, center keypoints often yield weak convolutional responses for objects with large homogeneous regions or hollow structures and for small objects, increasing the likelihood of missed detections and false alarms. To address these limitations, we propose LCMCA-Net, which integrates a large-kernel contour attention (LC) module and a multiscale corner alignment (MCA) module. The LC module uses large-kernel separable convolutions with stage-specific kernel sizes to enlarge the receptive field and capture fine-grained contour cues. The MCA module further uses parallel dilated convolutions to predict spatial offsets and modulation masks from center features, after which deformable sampling dynamically aligns the top-left and bottom-right features. The resulting geometry-related corner information is incorporated into the center branch. This integration compensates for weak center responses and improves consistency between the center and corner representations. Extensive experiments on the two public datasets, DIOR and NWPU VHR-10, show that LCMCA-Net attains state-of-the-art performance among the methods evaluated on these datasets. The source code is available at https://github.com/aicat777/LCMCA-Net