<p>Underwater target detection is crucial for ocean exploration, but existing methods struggle to achieve satisfactory results due to the complexity of the underwater environment. To enhance the accuracy and real-time performance of underwater detection models, we propose an improved YOLOv7 model. We introduce a multi-granularity feature attention method based on the Efficient Channel Attention (ECA) to help the model better adapt to the diverse conditions in the underwater environment, reducing focus on redundant information. Utilizing coordinate convolution provides the network with spatial awareness of input image coordinates, enabling more effective localization of target objects and reducing interference from similar background elements. To accommodate the features of small and dense underwater targets, we use normalized Wasserstein distance to measure the similarity of bounding boxes. On the Underwater Robot Picking Contest 2019 (URPC 2019) dataset, the mean Average Precision (mAP) of our improved network has reached 86.19%, which represents a 1.57% increase compared to the original YOLOv7 network. Additionally, the frames per second (fps) has achieved 124, surpassing the performance of the original network. This improvement is significantly superior to conventional target detection models, providing a faster and more accurate advantage for underwater target detection tasks in complex underwater environments.</p>