Algorithm 1: Adaptive Spatial and Channel Feature Fusion

Input: Feature map $F_I\in {\mathbb{R}}^{N\times C^{*}\times H\times W}$, $F_{\mathrm{concat}}\in {\mathbb{R}}^{N\times (C+C^{*})\times H\times W}$

Output: Fused feature $F^{IP}\in {\mathbb{R}}^{1\times C\times H\times W}$

AFSM: Channel Attention Branch

1. $h_{\mathrm{avg}}\leftarrow \mathrm{GAP}\left(F_I\right)$;

2. $h_{\max }\leftarrow \mathrm{GMP}\left(F_I\right)$;

3. $F_h\leftarrow \mathrm{Concat}(h_{\mathrm{avg}},h_{\max })$;

4. $F_h^{*}\leftarrow \sigma \left(\mathrm{ReLU}\left(F_h\right)\right)$;

5. $F_I^{*}\in {\mathbb{R}}^{N\times C\times H\times W}\leftarrow F_I\odot F_h^{*}$;

6. $F_I^{*}\in {\mathbb{R}}^{1\times C\times H\times W}\leftarrow \mathrm{ReLU}\left(\mathrm{Conv}3\mathrm{D}\left(F_I^{*}\right)\right)$;

DWLM: Spatial Attention Branch

7. $s\leftarrow \mathrm{GAP}\left(F_{\mathrm{concat}}\right)$;

8. $x\leftarrow \mathrm{ReLU}\left({\mathrm{FC}}_1\left(s\right)\right)$;

9. $w\leftarrow \mathrm{Softmax}\left({\mathrm{FC}}_2\left(x\right)\right)$;

10. $w\leftarrow \mathrm{reshape}(w,N,C,1,1)$;

11. $F_{\mathrm{concat}}^{*}\in {\mathbb{R}}^{N\times C\times H\times W}\leftarrow F_{\mathrm{concat}}\odot w$;

Feature Fusion

12. $F^{*}\leftarrow F_I^{*}\odot F_{\mathrm{concat}}^{*}$;

13. $F^{IP}\leftarrow \mathrm{ReLU}\left(\mathrm{Conv}2\mathrm{D}\left(F^{*}\right)\right)$;

return $F^{IP}$;