Fig. 1. Structure diagram of ACC residual feature enhancement module

Fig. 2. Schematic diagram of obtaining spatial features of receptive domain

Fig. 3. Schematic diagram of DSConv quantization process structure

Fig. 4. Changes on the x- and y- axis

Fig. 5. Structure diagram of BiFormer and BiFormer block models

Fig. 6. BRA module structure diagram

Fig. 7. Improved YOLOv8s network structure

Fig. 8. Comparison of detection results after adding convolutional modules. (a) YOLOv8s; (b) YOLOv8s+ODConv; (c) YOLOv8s+DCN; (d) YOLOv8s+SPD-Conv; (e) YOLOv8s+DSConv; (f) original image of SSDD dataset

Fig. 9. Comparison of detection results after adding different attention mechanisms. (a) YOLOv8s; (b) YOLOv8s+SimAM-attention; (c) YOLOv8s+CBAM-attention; (d) YOLOv8s+Self-attention; (e) YOLOv8s+CSWin-attention; (f) YOLOv8s+GLA-attention; (g) YOLOv8s+SU-attention; (h) YOLOv8s+BiFormer-attention; (i) original image of SSDD dataset

Fig. 10. Precision variation curve

Fig. 11. Images in SSDD dataset. (a) Single small target; (b) multiple small targets; (c) normal size target; (d) targets in a complex context

Fig. 12. Detection performance of YOLOv8s on SSDD dataset. (a) Single small target; (b) multiple small targets; (c) normal size target; (d) targets in a complex context

Fig. 13. Detection performance of proposed algorithm on SSDD dataset. (a) Single small target; (b) multiple small targets; (c) normal size target; (d) targets in a complex context

Fig. 14. Images in HRSID dataset. (a) Single small target; (b) multiple small targets; (c) multiple small targets in a complex context

Fig. 15. Detection performance of YOLOv8s on the HRSID dataset. (a) Single small target; (b) multiple small targets; (c) multiple small targets in a complex context

Fig. 16. Detection performance of our algorithm on the HRSID dataset. (a) Single small target; (b) multiple small targets; (c) multiple small targets in a complex context