<p>Underwater object detection (UOD) plays a crucial role in marine resource exploration and environmental conservation. To address the performance degradation caused by underwater image deterioration, several methods have been proposed to enhance image quality via underwater image enhancement (UIE) techniques to assist detection tasks. However, existing approaches often rely solely on detection loss for joint learning of UIE and UOD, limiting their effectiveness in improving detection accuracy. To overcome this limitation, we propose a dual supervision guided network (DSG-Net) for underwater object detection, which optimizes an end-to-end underwater detection model through the joint constraints of both detection loss and enhancement loss. During the training phase, we design a joint constraint loss to guide the UIE module in globally enhancing degraded images while preserving texture features critical for the detector. In the inference phase, DSG-Net uses the lightweight UIE module (LUIEM) to generate enhanced images. Through targeted processing of global and local features, it addresses challenges such as color cast and blurriness in underwater degraded images. Subsequently, a dual-branch backbone network with a multi-scale attention fusion module (MSAFM) is utilized to extract features from the original image and the enhanced image and perform adaptive fusion. This design effectively retains the discriminative features while suppressing the noise introduced by the enhancement algorithm. Considering the real-time requirements of real-world UOD applications, the design of DSG-Net has carried out lightweight design for the model size and modules. Compared with other underwater target detection models, such as ERL-Net, the number of parameters and computational load have been reduced by more than 90%.</p>