Optics and Precision Engineering, Volume. 33, Issue 8, 1274(2025)
Polarized image feature fusion in power inspection
Figures & Tables(22)
Fig. 3. Architecture of proposed PDBFNet. （a） shows the structure of the whole algorithm. （b） and （c） give the details of the dense block and the channel attention block
Fig. 8. Fusion samples of comparison algorithms. The area in the box region is enlarged to display the details
Fig. 10. Example of target detection results using different fusion algorithms to generate graphs
Table 1. Comparison of similar datasets
View tableView in ArticleTable 1. Comparison of similar datasets
Image Quantity Resolution Type Content TNO 261 768×576 IR+RGB Military Scene M3FD 4 500 1 024×768 IR+RGB Road Scene RoadScene 221 irregular IR+RGB Road Scene PolarLITIS 2 569 500×500 Polar+RGB Road Scene IPF （ours） 3 000 1224×1024 Polar+RGB Electricity Scene
Table 2. Effect of
w on image fusion quality in SSIM loss functionsView tableView in ArticleTable 2. Effect of
w on image fusion quality in SSIM loss functionsWindow Size EN MG SD SSIM DoLP S0 5×5 7.045 9.202 50.573 0.382 0.419 7×7 6.685 11.069 35.712 0.521 0.330 9×9 6.993 9.299 49.130 0.373 0.413 11×11 7.104 8.370 57.981 0.292 0.409 13×13 7.018 10.324 48.249 0.429 0.394 15×15 7.007 9.267 51.329 0.355 0.422 17×17 7.049 9.139 49.762 0.367 0.428
Table 3. Effect of multi-scale
w on image fusion quality in MASSIMView tableView in ArticleTable 3. Effect of multi-scale
w on image fusion quality in MASSIMMethod EN MG SD SSIM-DoLP SSIM-S0 MASSIM-Ⅰ 7.038 8.569 50.223 0.345 0.426 MASSIM-Ⅱ 7.155 9.547 55.968 0.355 0.411 MASSIM-Ⅲ 6.902 9.108 46.199 0.381 0.432
Table 4. Objective quality evaluation of raw images from IPF
View tableView in ArticleTable 4. Objective quality evaluation of raw images from IPF
Image EN SD MG S0 6.143 59.397 25.162 DoLP 5.771 32.919 72.400
Table 5. Objective quality evaluation of fusion model on IPF
View tableView in ArticleTable 5. Objective quality evaluation of fusion model on IPF
NO Fusion method EN SD MG MI QAB/F SSIM 1 NSCT 6.920 67.307 82.080 1.826 0.401 0.513 CVT 6.863 58.376 76.557 1.830 0.392 0.358 2 DIFNet 5.760 20.828 46.199 1.167 0.383 0.466 DDcGAN 5.630 23.311 33.996 1.098 0.388 0.438 PFNet 7.116 59.395 74.625 1.769 0.398 0.492 PDBFNet 7.206 76.917 66.170 2.744 0.401 0.469
Table 6. Objective quality evaluation of the PolarLITIS
View tableView in ArticleTable 6. Objective quality evaluation of the PolarLITIS
Image EN SD MG S0 6.562 62.292 25.827 DoLP 3.363 40.420 34.619
Table 7. Objective quality evaluation of fusion model on PolarLITIS
View tableView in ArticleTable 7. Objective quality evaluation of fusion model on PolarLITIS
NO Fusion method EN SD MG MI QAB/F SSIM 1 NSCT 6.765 54.187 44.853 4.055 0.393 0.518 2 DIFNet 5.673 20.743 16.435 3.214 0.381 0.472 DDcGAN 5.751 22.706 23.424 1.831 0.371 0.366 PFNet 6.707 41.557 39.686 3.572 0.382 0.534 PDBFNet 6.787 52.328 39.024 3.780 0.383 0.537
Table 8. Results of structure ablation experiment
View tableView in ArticleTable 8. Results of structure ablation experiment
Group Description EN MG SD SSIM DoLP S0 A Point Plus 7.127 9.359 53.189 0.337 0.219 Concat 7.114 8.837 51.813 0.334 0.225 Only-Spatial Fusion 7.082 9.170 51.684 0.350 0.226 Only-Channel Fusion 7.005 9.212 49.755 0.347 0.223 B No SE 7.151 9.307 54.208 0.342 0.225 PDBFNet（ours） 7.155 9.547 55.968 0.341 0.231
Table 9. Results of Loss function ablation experiment
View tableView in ArticleTable 9. Results of Loss function ablation experiment
EN MG SD SSIM DoLP S0 Single Average 7.048 9.025 52.652 0.396 0.331 Win-sets Average 7.227 9.590 57.770 0.396 0.360 Single Weight 7.196 10.061 57.942 0.354 0.344 Win-sets Weight 7.248 10.118 55.158 0.442 0.388
Table 10. Object detection results of comparative periment
View tableView in ArticleTable 10. Object detection results of comparative periment
Network Params/M Time/ms Precision Recall mAP 0.5 Yolov5-n 1.9 10.5 0.838 0.832 0.874 Yolov5-s 7.2 22.3 0.891 0.854 0.902 Yolov5-m 21.2 50.8 0.915 0.861 0.911 Yolov8-n 3.2 8.7 0.873 0.826 0.877 Yolov8-s 11.2 26.2 0.878 0.85 0.882 Yolov11-s 9.4 26 0.886 0.874 0.903
Table 11. Target detection results using raw images
View tableView in ArticleTable 11. Target detection results using raw images
Image mAP 0.5 Precision Recall S0 0.891 0.854 0.902 DoLP 0.852 0.843 0.873
Table 12. Target detection results using fused images
View tableView in ArticleTable 12. Target detection results using fused images
Fusion method mAP 0.5 Precision Recall NSCT 0.861 0.866 0.898 CVT 0.851 0.871 0.887 DIFNet 0.858 0.848 0.885 DDcGAN 0.874 0.852 0.889 PFNet 0.852 0.812 0.863 PDBFNet 0.915 0.873 0.916
Tools
Get Citation
Copy Citation Text
Mengyao NI, Yuanlong PENG, Shang HU, Longchuan YAN, Jinkun ZHENG, Danhua CAO. Polarized image feature fusion in power inspection[J]. Optics and Precision Engineering, 2025, 33(8): 1274
Paper Information
Category:
Received: Nov. 8, 2024
Accepted: --
Published Online: Jul. 1, 2025
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20