As the exploration of marine resources goes deeper, underwater object detection plays an increasingly important role. However, underwater objects feature various size which causes great confusion for detectors. Additionally, underwater images often encounter degradation problems, such as blur, low contrast, color deviation and uneven illumination. These low-quality underwater images severely affect detection performance. In this paper, we propose an MDI-Net to improve the detection performance effectively for underwater detectors. We propose a Multi-level detection layers interaction module to enhance information interaction among detection layers. It comprehensively considers the information of all detection layers and then uses spatial attention to force each detection layer to pay more attention on the most relevant regions, thereby improving the recognition ability of underwater objects. Then we design a novel image-level retraining strategy suitable for one-stage detectors to mine difficult underwater images that are hard for detection during training. Therefore, the detectors can learn more information and extract finer features from these hard samples to enhance their adaptability to the underwater environment. The extensive experiments on DUO datasets demonstrate the favorable performance of the proposed MDI-Net against the baseline and other object detectors in terms of detection accuracy.