Seabed pockmarks are important geomorphological indicators of shallow fluid migration, shallow gas activity, and gas hydrate-related processes. However, automatic pockmark detection from shaded-relief bathymetric imagery remains challenging because many pockmarks have small image-space footprints, occur densely, have weak boundaries, and are embedded in complex seabed backgrounds. To address these issues, this study proposes FSLA-RTDETR, an improved four-scale RT-DETR network for dense small-object seabed pockmark detection. The model introduces SAC-C2f to enhance high-level feature representation, the LPEA Encoder to improve global contextual modeling, and a DGCST-based bidirectional feature fusion network to integrate shallow spatial details and deep semantic information. In addition, the high-resolution feature is incorporated into the decoder to improve small-object localization. A pockmark detection dataset was constructed from publicly available BOEM bathymetric data from the northern Gulf of Mexico, containing 880 images and 8013 annotated pockmark instances. Experimental results show that FSLA-RTDETR achieves mAP50, mAP75, and mAP50–95 values of 0.9218, 0.6221, and 0.5485, respectively. Compared with the baseline RT-DETR, mAP75 and mAP50–95 increase by 6.01 and 2.22 percentage points, respectively. Under the present no-pretraining setting, comparative experiments further demonstrate that FSLA-RTDETR achieves superior high-IoU localization performance compared with the representative object detection models evaluated in this study, providing an effective method for automatic seabed pockmark detection and geomorphological mapping.