In response to issues such as the low recognition accuracy of small objects and low-resolution images in traditional diver machine recognition algorithms, a underwater diver recognition algorithm which is named IM-YOLOv9 based on improved YOLOv9 (You Only Look Once version 9) is proposed to enhance the precision and reliability of diver recognition for applications in marine resource development. The algorithm introduces SPD-Conv (Spatial Depth Conversion Convolution) for image processing, incorporates the SENetV2 (Squeeze and Excitation Networks Version 2) module, adds the MPDIoU (Multi-Path Distance Intersection over Union) loss function, and includes the SCINet (Self-Calibrating Illumination Network) module. These modifications aim to improve recognition accuracy without affecting the model’s operating speed. Results indicate that the improved IM-YOLOv9 model outperforms the original YOLOv9 model, with a 3.88% increase in precision, a 4.35% increase in recall, a 2.56% improvement in the F1 score, and a 4.76% increase in mean average precision (mAP). These enhancements make the improved model better suited for underwater diver recognition.