Aiming at the problems of seal missing in files of the dossiers. We analyze the object detection under complex background, which is an important checkpoint in dossiers review. The seals on the files are usually unclear, such as blurry, incomplete, overlapping, which are difficult to recognize by the model. In order to solve this problem, we propose a seal detection model, which is based on dilated convolution neural network. More specifically, focus on the clustering algorithm to get the anchors, which are more suitable for seal object. In order to increase the performance of the model in complex background, we use dilated convolutions in the first three sampling layer and ASFF in the neck network. Finally, we use GIOU loss in the loss function of YOLOv3. The F1 value of the seal detection model increased by 8.43 percent, which reaches the seal detection F1 index.