Pretrained Detail Enhancement Framework for Remote Sensing Object Detection

Remote sensing object detection faces significant challenges due to the varying resolutions of images and the diverse shapes of objects. In this paper, we introduce an innovative approach to enhance object detection networks in remote sensing images by addressing these issues through a novel two-stage training process. Our key contributions include the development of a combined low-resolution pretraining and high-resolution fine-tuning strategy, which significantly improves the model’s generalization capability across different resolutions. We have designed a Pretrained Detail Enhancement and Restoration (PDER) framework that enhances feature extraction from low-resolution images, thereby boosting the baseline network’s ability to handle blurred and intricate objects. Additionally, we propose an Energy-Efficient Shape Adaptive Convolution (EESA) that integrates a depthwise separable mechanism with shape-adaptive feature maps, enabling efficient capture of complex object edges without substantial computational overhead. Compared to state-of-the-art methods such as YOLOv7, our model achieves an absolute average precision improvement of 2.34% and 2.57% on the DIOR and xView datasets, respectively. Extensive experiments on multiple remote sensing datasets demonstrate our approach’s effectiveness, achieving significant improvements in detection accuracy and computational efficiency, thus highlighting its potential for practical applications.