Maintaining and monitoring underwater infrastructure, key to efficient and low-risk asset management, relies heavily on underwater object detection (UOD). Sonar, favored in murky, low-light underwater conditions, faces challenges like low resolution and poor contrast in its images, affecting object detection accuracy. This paper introduces a novel deep learning approach for UOD using multibeam forward-looking sonar data, leveraging the YOLOv7 (abbreviation of You Only Look Once version 7) architecture. It includes tailored improvements in data preprocessing, feature fusion, and loss functions, outperforming existing sonar methods in object classification. Tested on an underwater remotely operated vehicle (ROV), the proposed framework shows enhanced target classification and transfer learning, suggesting potential for broad application in underwater structural monitoring and autonomous management.