Underwater object detection based on forward-looking sonar (FLS) images has been a popular topic in recent years and numerous detection paradigms designed and optimized for optical images have been applied to FLS images. However, most of these detection paradigms are hard to achieve top performance. In this paper, we propose the Cross Stage Partial Block (DCSPBlock) to build backbone network, which is called DCSP-chain. Specifically, DCSPBlock consists of Res(1) Block and skip-layer with dilated convolution (SLDC). Therefore, the DCSP-chain can reduce redundant gradient information and provide effective receptive fields. Additionally, the BiFPN with 4 levels serves as multi-scale feature fusion network to increase FLS detection accuracy. Benefiting from DCSP-chain and multi-scale feature fusion, the proposed DCSP-YOLOv5 paradigm significantly improves the perception ability of objects. The heat map further visualizes the attention of DCSP-YOLOv5 to FLS image. Furthermore, numerous experiments show that the proposed paradigm achieves competitive performance on URPC2021 FLS dataset.