Composite Perception and Multiscale Fusion Network for Arbitrary-Oriented Object Detection in Remote Sensing Imagery

Arbitrary-oriented object detection has been widely used in remote sensing image object detection, mainly benefiting from accurate boundary boxes with less background information. However, the fixed receptive field feature extraction adopted by current mainstream methods limits the dynamic focusing ability of the model, resulting in insufficient context information capture, which leads to small object submergence. Furthermore, the complex background and dense distribution of objects in remote sensing images also increase the difficulty of detection, making it vulnerable to noise and fuzzy boundary interference. Therefore, we propose a new arbitrary-oriented object detection (AOOD) network called composite perception and multiscale fusion network (CPMFNet). First, to aggregate different levels of context information of the object and effectively enhance the context modeling ability of the detector, we propose an adaptive spatial perception module (ASPM) and insert it into the residual network as a plug-and-play module to realize dynamic receptive field convolution. Second, we redesigned a feature pyramid network (FPN) equipped with a composite perception and context integration module (CPCIM) and detail noise balance filter (DNBF). CPCIM realizes cross-scale feature fusion, mines position and detail features of shallow networks, enhances global semantic expression, and weakens interference of background information. DNBF can effectively suppress the interference of noise and fuzzy boundaries and guide deep networks to alleviate the inundation of small objects. Experiments on four remote sensing benchmark datasets (DOTA-v1.0, DOTA-v1.5, DIOR-R, and HRSC2016) show that CPMFNet (ours) achieves SOTA performance, in which the mean average precision (mAP) is 78.23% on the DOTA-v1.0 dataset, 70.85% on the DOTA-v1.5 dataset, and 67.58% and 98.61% on the DIOR-R and HRSC2016 datasets, respectively.