<p>Small object detection is a challenging issue in computer vision. Most of the existing methods measure the similarity between two bounding boxes using the intersection over union. However, intersection over union is very sensitive to small shifts between two bounding boxes, which is not conducive to small object detection. In common decoupled-head structure, the design that classification and localization enjoy the shared input cannot solve the conflict between localization and classification well, resulting in an imperfect balance between these two tasks. To overcome the above limitations, we first propose the center deviation degree. By integrating center deviation degree with the intersection over union variants, we establish a novel hybrid evaluation metric that substantially improves detection accuracy across multiple object scales, while specifically addressing the critical challenges associated with small object detection. Moreover, we generate semantic context-rich features and detailed spatial features as the inputs for classification and localization branch, respectively. Decouple the two tasks by providing them with independent input features for more accurate object detection. Extensive experiments show that, when equipped with hybrid evaluation metric and classification-localization context decoupling, our method achieves gains of 1.8, 1.7, 1.7 and 1.9 points on very tiny objects, tiny objects, small objects and medium objects respectively, compared with the baseline.</p>