Recently, remote sensing object detection has gained much attention in computer vision. Oriented object detection faces additional challenges due to the inherent differences between remote sensing images and those from natural scenes, such as detecting objects with different directions and scales. Recent methods improve the model structure of feature pyramid network (FPN) to solve scale problems, but most of them disregard the default loss of channel information during lateral connections and the semantic gap among different levels of feature maps during fusion. In this paper, we propose a context-enhanced FPN (CEFPN) consisting of an information compensation module and an elaborate fusion to address the aforementioned problems. Specifically, we introduce the PixelShuffle to address the problems of channel information loss and design a novel fusion module to enhance contextual features. In addition, we propose a shape-aware soft label assignment (SSLA) method to alleviate the sensitivity of assignment results to scale variation and angle variation. We design a scoring function to compute a quality score for each anchor box and dynamically select high-quality anchor boxes as positive samples. Moreover, the quality score is utilized to weigh the training samples in the loss function, prioritizing high-quality anchors during the learning process. Experimental results on DOTA and HRSC2016 datasets demonstrate the effectiveness of our proposed method.