Abstract: The detection accuracy of the state-of-the-art deep learning object detection algorithms for small objects is notably lower than that of medium and large objects. In order to solve this problem, we propose a self-supervised small object detection method by combining self-supervised learning and the task of small object detection. Firstly, a classification network is pre-trained by a self-supervised contrastive learning algorithm and then transferred to the small object detection model. Compared with the traditional supervised pre-training, the self-supervised pre-training can extract universal feature representations from massive unlabeled data, significantly reduce the cost of data annotation, and enhance the generalization ability of the model in downstream tasks. Secondly, in order to improve the ability of the model to extract small object features, we adopt a contrast learning method based on pixel-level classification. Compared to the classic image-level contrastive learning, pixel-level contrast pays more attention to the details in the image, the backbone network has stronger feature representation ability that achieves better performance when transferred to small object detection tasks. Meanwhile, we optimize and improve the structure of the detection network by adding synchronized batch normalization layer to ensure that the distribution of input data does not deviate due to parameter updating. Finally, we discuss the model training methods based on full fine-tuning and partial fine-tuning, conducting experiments under conditions where different layers of the backbone network are frozen. The results show that our method outperforms traditional supervised pre-training and classic self-supervised methods on small object detection tasks, and the average detection precision on SODA datasets is improved by 1.3% and 1.6%, respectively.