Fig. 1. Three main surgical instruments in RET1 dataset. (a) Vitrectomy cutter; (b) Internal limiting membrane peeling forceps; (c) Light pipe

Fig. 2. Overall network architecture of the proposed RM-YOLO model

Fig. 3. Cascading of three attention mechanisms in the detection head

Fig. 4. Integrated structure of the proposed dynamic head and decoupled head

Fig. 5. Diagram of the SGConv convolution layer

Fig. 6. RSGCSP feature extraction modules

Fig. 7. Transformation process of the reparameterization part

Fig. 8. Schematic diagram of proposed DeltaIoU

Fig. 9. Results of the loss function experiments. (a) mAP at different values of α; (b) Bounding box regression loss corresponding to different IoUs; (c) mAP corresponding to different IoUs

Fig. 10. The results of three main types of surgical instruments in the experiment

Fig. 11. Detection results of different algorithms for a single surgical instrument in complex scenarios. (a) Severe instrument reflection; (b) Severe distortion in microscope imaging

Fig. 12. Detection results of different algorithms for multiple surgical instruments in complex scenarios. (a) High instrument similarity; (b) Microscope out of focus; (c) Occlusion between instruments

Fig. 13. Detection of surgical instruments during vitrectomy under different lighting conditions by the proposed algorithm