Remote Sensing Small Object Detection Based on Cross-Layer Attention Enhancement

Fig. 5. Overall structure of the cross-layer attention. (a) Complete flow of the cross-layer attention; (b) catt module of the cross-layer attention; (c) satt module of the cross-layer attention

Download full size

Fig. 6. Distribution statistics of label boxes in the data set. (a) Distribution of the size of the label box; (b) distribution of the center points of the label box

Download full size

Fig. 7. Effect analysis on hyperparameter $γ$ . (a) Influence of hyperparameter $γ$ on detection performance for small object; (b) influence of hyperparameter $γ$ on detection performance for medium object; (c) influence of hyperparameter $γ$ on detection performance for large object; (d) influence of hyperparameter $γ$ on detection performance for object

Download full size

Fig. 8. Heat map generated by the proposed CLAT module by using the Grad-CAM method

Download full size

Fig. 9. Effect contrast of the proposed model and baseline on the test set. (a) baseline; (b) proposed model

Download full size

Table 1. Contrast experiment

View table

Table 1. Contrast experiment

Method	Backbone	APs /%	mAP /%
SSD	VGG16		58.6
YOLOv3	Darknet-53	11.6	57.1
Faster R-CNN with FPN	ResNet‐101		65.1
Mask R-CNN with FPN	ResNet‐101		65.2
Libra R-CNN	ResNet‐101	14.9	79.7
Dynamic R-CNN	ResNet50	12.1	77.3
YOLOx	Darknet-53	17.3	85.7
YOLOv5x6	CSPDark-53	19.9	86.6
Proposed method	CSPDark-53	23.4	86.4

Table 2. Ablation experiment
View table
Table 2. Ablation experiment
Baseline RCP CLAT mAP /% APs /% APm /% APl /%
√ 85.5 17.5 52.0 75.6
√ √ 85.4 21.6 51.7 75.4
√ √ 86.2 19.2 52.2 76.1
√ √ √ 86.4 23.4 52.3 75.9

Table 3. Sample distribution of the DIOR dataset

View table

Table 3. Sample distribution of the DIOR dataset

Class	Number of samples	Class	Number of samples	Class	Number of samples	Class	Number of samples
golf-field	881	dam	824	stadium	1003	airport	1071
train-station	811	chimney	1340	harbor	4447	bridge	3161
expressway-service-area	1743	basketball-court	2658	overpass	2478	airplane	8100
expressway-toll-station	1028	ground-track-field	2390	windmill	4371	vehicle	32180
baseball-field	4674	tennis-court	9621	storage-tank	20717	ship	50207

Table 4. Time-consuming comparison
View table
Table 4. Time-consuming comparison
Model Backbone Time /s
baseline CSPDark-53 0.06
Proposed model CSPDark-53 0.09

Tools

Get Citation

Copy Citation Text

Xingbo Han, Fan Li. Remote Sensing Small Object Detection Based on Cross-Layer Attention Enhancement[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1228011

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites