Adaptive Circular Receptive Field for Remote Sensing Image Target Detection

Remote sensing images contain a variety of targets. The circular and square-like targets are very common. However, general methods are prone to incorporating background information and have problems with rotation angles when detecting circular and square-like targets. The Adaptive Circular Receptive Field Network (ACRFNet) is proposed to address these problems. Firstly, a radius detection head based on the center point is designed, which predicts the center point and radius of the target to obtain a bounding circle. Compared with the rectangular bounding box, the bounding circle contains less background information and has rotation invariance, achieving in more accurate localization. Secondly, in order to better extract the features of circular targets and cope with scale changes, a multi-scale feature fusion method based on adaptive circular convolution is proposed. Adaptive circular receptive fields replace the ordinary convolution of square receptive fields, and a feature pyramid structure is introduced to enrich the semantic information of the feature map. Finally, in order to reasonably evaluate the performance of the model, Circle IOU is used to calculate mAP to fit the bounding circle. The experimental results on the NWPU VHR-10 and DIOR datasets demonstrate that the model has significantly improved detection accuracy for circular and square-like targets and has better generalization for other types of targets.