Small Object Detection Based on Spatial-Frequency Separated Cross-Attention Swin Transformer

Keping Wang; Bingqian Suo; Gaopeng Zhang; Yi Yang; Wei Qian

doi:10.3788/LOP242531

Laser & Optoelectronics Progress, Volume. 62, Issue 16, 1628001(2025)

Small Object Detection Based on Spatial-Frequency Separated Cross-Attention Swin Transformer

Keping Wang^1,2, Bingqian Suo^1,2、*, Gaopeng Zhang³, Yi Yang^1,2, and Wei Qian¹

Author Affiliations

¹School of Electrical Engineering and Automation, Henan Polytechnic University, Jiaozuo 454003, Henan , China

²Henan International Joint Laboratory of Direct Drive and Control of Intelligent Equipment, Henan Polytechnic University, Jiaozuo 454003, Henan , China

³Xi'an Institute of Optics and Precision Mechanics of CAS, Xi'an 710119, Shaanxi , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Structure of SF-SCST

Download full size

Fig. 2. Structure of stage 1 in SF-SCST network

Download full size

Fig. 3. Structure of SFDF module

Download full size

Fig. 4. Structure of CS-swin module

Download full size

Fig. 5. Comparison of detection results for small targets in simple environment

Download full size

Fig. 6. Comparison of detection results for small targets in complex environment

Download full size

Fig. 7. Comparison of detection results for small targets in dark environment

Download full size

Fig. 8. Heat map comparison of different algorithms

Download full size

Fig. 9. Comparison of model accuracy and loss between SF-SCST and Swin Transformer algorithms. (a) Model accuracy; (b) classification loss; (c) regression loss

Download full size

Fig. 10. Visualization of the module outputs for SF-SCST and Swin Transformer algorithms

Download full size

Fig. 11. Accuracies of SF-SCST algorithm for small objects with different categories and scales on two datasets. (a) different categories; (b) different scales

Download full size

Table 1. Comparison experimental results of different algorithms on two datasets

View table

Table 1. Comparison experimental results of different algorithms on two datasets

Algorithm	UAV-DA dataset				VisDrone dataset				GFLOPs	Params /10⁶
Algorithm	mAP /%	mAP₅₀ /%	mAP_S /%	mAP_M /%	mAP /%	mAP₅₀ /%	mAP_S /%	mAP_M /%	GFLOPs	Params /10⁶
IterDet	24.4	65.7	17.0	23.9	26.3	45.7	17.0	38.0	23.36	41.36
RetinaNet	23.7	62.1	14.2	23.8	21.4	36.2	11.2	34.2	10.07	36.13
GCNet	24.7	65.8	15.7	27.2	25.6	43.0	16.1	37.4	23.37	51.13
GFL	26.0	68.7	17.9	28.9	24.6	42.6	15.5	35.7	10.04	32.04
TridentNet	23.4	61.9	15.1	22.7	20.6	36.5	11.8	30.5	738.17	32.77
Swin Transformer	25.2	67.7	17.2	23.2	27.3	45.2	17.6	39.7	23.39	44.79
Swin+WDFC	25.7	69.1	18.2	27.7	28.6	47.1	18.8	41.5	24.01	45.19
Swin+SFDF	25.6	69.3	17.6	29.2	27.9	45.9	17.9	40.6	24.73	48.64
SF-SCST （proposed）	25.5	69.3	18.1	30.2	28.1	46.0	18.3	40.6	26.68	81.91

Table 2. Comparison experimental results between SF-SCST and Swin Transformer algorithms on the Wider Person dataset
View table
Table 2. Comparison experimental results between SF-SCST and Swin Transformer algorithms on the Wider Person dataset
Algorithm mAP mAP₅₀ mAP_S mAP_L
Swin Transformer 12.3 24.8 5.8 18.7
SF-SCST （proposed） 12.3 24.8 6.1 19.1

Table 3. Results of ablation experiments on different datasets

View table

Table 3. Results of ablation experiments on different datasets

WDFC	SFDF	CS-swin	UAV-DA dataset				VisDrone dataset
WDFC	SFDF	CS-swin	mAP	mAP₅₀	mAP_S	mAP_M	mAP	mAP₅₀	mAP_S	mAP_M
—	—	—	25.2	67.7	17.2	23.2	27.3	45.2	17.6	39.7
√	—	—	25.7	69.1	18.2	27.7	28.6	47.1	18.8	41.5
—	√	—	25.6	69.3	17.6	29.2	27.9	45.9	17.9	40.6
—	—	√	26.4	69.8	17.6	30.6	28.5	46.7	18.8	41.1
√	—	√	26.1	69.5	18.7	30.1	28.3	46.6	18.9	40.5
√	√	—	25.7	69.9	18.6	29.7	27.9	46.2	18.4	40.1
—	√	√	25.6	69.4	18.0	29.1	27.2	45.0	17.5	39.5
√	√	√	25.5	69.3	18.1	30.2	28.1	46.0	18.3	40.6

Table 4. Comparison experimental results of WDFC module implementation

View table

Table 4. Comparison experimental results of WDFC module implementation

Algorithm	UAV-DA dataset				VisDrone dataset
Algorithm	mAP	mAP₅₀	mAP_S	mAP_M	mAP	mAP₅₀	mAP_S	mAP_M
Swin Transformer	25.2	67.7	17.2	23.2	27.3	45.2	17.6	39.7
+ WDFC（2×2）	25.7	69.1	18.2	27.7	28.6	47.1	18.8	41.5
+ WDFC（1×4）	26.0	69.9	18.0	27.6	28.4	47.0	18.5	41.2

Tools

Get Citation

Copy Citation Text

Keping Wang, Bingqian Suo, Gaopeng Zhang, Yi Yang, Wei Qian. Small Object Detection Based on Spatial-Frequency Separated Cross-Attention Swin Transformer[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1628001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites