Remote Sensing Image Detection Based on Dense Connected Networks

Table 1. Object size division standard
View table
Table 1. Object size division standard
Area (0,32²) [32²,96²] (96²,∞)
Classes small medium large
Percentage /% 28 32 40

Table 2. Effect of number of dense blocks on detection results under the same network depth
View table
Table 2. Effect of number of dense blocks on detection results under the same network depth
Num-block Large /% Medium /% Small /% mAP /% Time /s
3 84.58 80.28 78.43 81.48 0.014
4 83.67 79.34 75.90 80.11 0.010
5 84.25 57.41 50.65 66.25 0.008

Table 3. Experimental results obtained by changing the number of feature layers and network depth when number of dense block is 4
View table
Table 3. Experimental results obtained by changing the number of feature layers and network depth when number of dense block is 4
Growth-rate Depth Params /10⁶ Time /s mAP /%
12 40 0.1 0.010 80.11
12 100 0.6 0.021 85.99
24 40 0.6 0.020 85.24
24 100 2.4 0.070 86.59
40 190 22.6 0.234 87.72

Table 4. Comparison of detection results of different algorithms
View table
Table 4. Comparison of detection results of different algorithms
Algorithm X_TP X_FP X_FN P R
SSD 5211 943 527 0.90 0.85
YOLO v3 5612 694 376 0.93 0.89
Ours-40 5652 639 390 0.94 0.90
Ours-100 6013 409 259 0.96 0.93

Table 5. Improvement effect of network
View table
Table 5. Improvement effect of network
Algorithm mAP_large /% mAP_medium /% mAP_small /% mAP /% Time /s
Densenet-40 82.37 78.12 64.92 76.12 0.008
Ours-40 83.67 79.34 75.90 80.11 0.010

Tools

Get Citation

Copy Citation Text

Zexing Du, Jinyong Yin, Jian Yang. Remote Sensing Image Detection Based on Dense Connected Networks[J]. Laser & Optoelectronics Progress, 2019, 56(22): 222803

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites