Fig. 1. Multi-heterostructure filtering processing module

Fig. 2. Original images and filtering results of three channels under different backgrounds. (a) Scenario 1; (b) scenario 2; (c) scenario 3

Fig. 3. Schematic diagram of target location regression

Fig. 4. Schematic diagram of CBAM

Fig. 5. Comparison of curves of sigmoid function and fast_sigmoid function

Fig. 6. Structure diagram of proposed method

Fig. 7. Raw images in dataset. (a) Scenario 1; (b) scenario 2; (c) scenario 3; (d) scenario 4; (e) scenario 5; (f) scenario 6; (g) scenario 7; (h) scenario 8; (i) scenario 9

Fig. 8. Comparison of PR curves of different detection methods

Fig. 9. Changes of shallow feature maps before and after multi-heterostructure filtering. (a) Original images; (b) P3 feature maps before processing; (c) P3 feature maps after processing; (d) P4 feature maps before processing; (e) P4 feature maps after processing

Fig. 10. Effect of different anchor sizes on PR results

Fig. 11. Changes of shallow feature maps before and after using attention mechanism. (a) Original images; (b) P3 feature maps before processing; (c) P3 feature maps after processing; (d) P4 feature maps before processing; (e) P4 feature maps after processing

Fig. 12. Change curves of PR before and after using attention mechanism

Fig. 13. Change curves of PR before and after clipping