Remote sensing image analysis has extremely broad application prospects in the fields such as land and resources management and ocean monitoring. Deep learning technology has made breakthroughs in the field of image processing. However, due to the inherent characteristics such as large size and small and dense objects of remote sensing, the deep learning methods for common images has some problems such as inaccurate positioning, difficult small object detection, and low large image detection accuracy in object detection for remote sensing images. Aiming at the above problems, this paper proposes a new object detection algorithm DFS for remote sensing images. Compared with traditional machine learning methods, DFS optimizes and designs a new dimension clustering module, customized loss function and sliding window segmentation detection mechanism. Among them, dimension clustering module optimizes priori anchors by designing clustering mechanism to improve the positioning accuracy, the customized loss function improves detection accuracy of small objects such as ships, and the sliding window segmentation detection solves the problem of low detection accuracy of large images. The comparative experiments on the classical remote sensing datasets show that, compared with YOLOv2, DFS improves mAP by 25.6% and greatly improves the detection efficiency of small objects and large images.