In the realm of underwater object detection, challenging conditions, characterized by blurriness and complexity, suppress the representative power of general deep backbone features. In this work, we propose an innovative decoupled head structure building upon the YOLOv7 framework. This structure segregates classification and regression branches to better capture the semantic features required for each subtask. In order to make effective use of the contradiction between classification and regression, we introduce the adjacent feature layer as a complementary operator to harmonizing subtasks. To address the issue of image blurriness in the underwater environment, a probabilistic modeling approach was adopted for regression box handling. The ultimate detection outcome is determined jointly by the classification and regression branches, enhancing the overall consistency of the category and bounding box results. An additional branch is introduced within the classification branch and seamlessly integrated to further augment the coherence of the detection outcomes. This comprehensive approach effectively addresses the challenges posed by the underwater environment, significantly improving the accuracy and robustness of underwater object detection.