Abstract: Underwater images suffer from color distortion and loss of details, which seriously affect the visual perception ability of underwater robots. To achieve image enhancement while improving detection accuracy, a multi-task learning framework is proposed for underwater image enhancement and object detection based on contrastive learning, which not only generates visually friendly images, but also improves object detection accuracy, achieving image enhancement for object detection tasks. To address the issue of unclear target texture features, a region generation module for detection tasks is used to construct positive and negative image blocks for contrastive learning, ensuring that the target region is closer to the original image in the feature space. Moreover, the detected gradient information is used to guide image enhancement in a direction beneficial for target detection. Additionally, an image translation method based on the cycle-generative adversarial network is proposed to learn and preserve clear image features for image enhancement, eliminating the need for paired underwater images and reducing data requirements. Finally, the enhancement algorithm is validated on the EUVP, U45, and UIEB datasets, and the detection algorithm is validated on the RUOD, URPC2020, and RUIE datasets. The experimental results show that the proposed algorithm can effectively correct color distortion in subjective vision, while preserving the structural texture of the original image and the target; in terms of objective indicators, the peak signal-to-noise ratio reaches 24.57 dB and the structural similarity reaches 0.88. The detection accuracy is improved by an average of 2% on Faster R-CNN (region-based convolutional neural network) and YOLOv7 (you only look once, version 7) algorithms after image enhancement.