Feature pyramid fully convolutional network for small and medium object detection in high-resolution remote sensing images

In recent years, deep neural networks have demonstrated significant success in object detection. However, detecting small and medium-sized objects in large-scale, high-resolution remote sensing images (HRSI) remains a challenging task with important military and civilian applications. To address this challenge, we propose an end-to-end feature pyramid fully convolutional network (FP-FCN). Our approach is informed by a detailed analysis of the differences between HRSI and natural images, particularly in terms of color, texture, shape, and spatial layout characteristics. This analysis guided the design of the FP-FCN framework. First, to accommodate objects at varying scales, we employ a feature pyramid network (FPN) to generate multi-level feature maps. Second, considering the high speed and precision of the region-based fully convolutional network (R-FCN) in object detection tasks, we adopted it as the base structure and implemented two parallel operations of location-sensitive pooling on the generated feature maps. Classification results are extracted from multi-level feature maps, whereas regression results are obtained from large-scale feature maps. Third, to mitigate the potential reduction in detection efficiency introduced by multi-level feature maps, we adjusted the number of channels in the feature map according to the number of object types relevant to the HRSI detection task. Finally, experimental results on the PASCAL VOC dataset demonstrate the effectiveness of our proposed method. For the most difficult-to-detect small objects, FP-FCN improves the mean average precision (mAP) by 3.2%, 6.6%, and 5.2% compared with Faster R-CNN + FPN, YOLOF, and R-FCN, respectively. In the HRSI experiment, we developed a small dataset targeting the core components of nuclear power plants as detection objects. When compared with existing object detection networks, our method exhibits clear advantages.