X-Ray Rapid Decentralized Detection Model of U<sup>2</sup>-Net with Dilated Convolution Optimization

Jiao Wang; Meng Wu; Jiankai Xiang

doi:10.3788/LOP241315

Laser & Optoelectronics Progress, Volume. 61, Issue 15, 1512006(2024)

X-Ray Rapid Decentralized Detection Model of U²-Net with Dilated Convolution Optimization

Jiao Wang¹, Meng Wu^2、*, and Jiankai Xiang³

Author Affiliations

¹School of Mathematics and Computer Science, Yan'an University, Yan'an 716000, Shaanxi, China

²School of Information and Control Engineering, Xi'an University of Architecture and Technology, Xi'an 710055, Shaanxi, China

³Shaanxi Institute for the Preservation of Cultural Heritage, Xi'an 710075, Shaanxi, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Shuffle Attention model structure diagram

Download full size

Fig. 2. Structure diagram of UD-block model

Download full size

Fig. 3. Details of the 6-layer UD-block module

Download full size

Fig. 4. Schematic diagram of misplaced links

Download full size

Fig. 5. Pyramid-divided attention module network framework

Download full size

Fig. 6. Improved image segmentation network framework of U²-Net

Download full size

Fig. 7. X-ray images of copper mirror under test (top is whole picture, bottom is partial picture on the lower left)

Download full size

Fig. 8. Segmentation and comparison of the copper mirror under test. (a) image of the copper mirror under test; (b) Manually annotated maps; (c) U²-Net; (d) U²-Net+PSA; (e) U²-Net+PSA+SA; (f) U²-Net+PSA+SA+ misplaced link; (g) U²-Net+PSA+SA+ misplaced link +UD-block

Download full size

Fig. 9. Comparison of the segmentation details of the copper mirror under test. (a) Image of the copper mirror under test; (b) manually annotated maps; (c) U²-Net; (d) U²-Net+PSA; (e) U²-Net+PSA+SA; (f) U²-Net+PSA+SA+ misplaced link; (g) U²-Net+PSA+SA+ misplaced link +UD-block

Download full size

Fig. 10. Comparison of algorithm segmentation of copper mirror under test

Download full size

Table 0. [in Chinese]

View table

Table 0. [in Chinese]

Group	Evaluation index	Dice	Accuracy	Jaccard	HD	PSNR	RMSE
Group 1	Fig. 8（c）	0.5179	0.5013	0.3495	238.0000	18.4814	30.3717
	Fig. 8（d）	0.5745	0.4927	0.4030	148.0540	18.6224	29.8827
	Fig. 8（e）	0.8256	0.7798	0.7030	196.4612	22.8313	18.4065
	Fig. 8（f）	0.8891	0.9306	0.8003	72.4983	26.3005	12.3456
	Fig. 8（g）	0.8975	0.9513	0.8141	34.7131	26.9814	11.4148
	Change	↑0.3796	↑0.4500	↑0.4646	↓203.2869	↑8.5000	↓18.9569
Group 2	Fig. 8（c）	0.7472	0.6213	0.5965	165.7227	19.8023	26.0871
	Fig. 8（d）	0.7447	0.6122	0.5933	136.0331	19.8380	25.9802
	Fig. 8（e）	0.8505	0.8201	0.7398	169.6850	19.5857	26.7457
	Fig. 8（f）	0.8816	0.9122	0.7883	127.3106	22.9702	18.1146
	Fig. 8（g）	0.8929	0.9432	0.8064	37.8550	23.7394	16.5794
	Change	↑0.1457	↑0.3219	↑0.2099	↓127.8677	↑3.9371	↓9.5077
Group 3	Fig. 8（c）	0.6916	0.5324	0.5286	105.2616	20.9815	22.7754
	Fig. 8（d）	0.7148	0.5610	0.5562	122.9675	21.1984	22.2137
	Fig. 8（e）	0.8778	0.9436	0.7822	100.4988	25.3134	13.8315
	Fig. 8（f）	0.8865	0.9481	0.7962	69.1158	26.3113	12.3304
	Fig. 8（g）	0.8890	0.9555	0.8001	69.1158	26.3602	12.2612
	Change	↑0.1974	↑0.4231	↑0.2715	↓36.1458	↑5.3787	↓10.5142
Group 4	Fig. 8（c）	0.8705	0.9369	0.7708	24.4131	23.9943	16.0999
	Fig. 8（d）	0.8705	0.9369	0.7708	24.4131	23.9943	16.0999
	Fig. 8（e）	0.8722	0.9475	0.7734	17.0000	29.0232	9.0236
	Fig. 8（f）	0.8728	0.9488	0.7743	17.0000	29.1287	8.9147
	Fig. 8（g）	0.8743	0.9545	0.7767	8.0623	30.3451	7.7497
	Change	↑0.0038	↑0.0176	↑0.0059	↓16.3508	↑6.3508	↓8.3502
Group 5	Fig. 8（c）	0.7694	0.8277	0.6253	138.7155	19.3674	27.4265
	Fig. 8（d）	0.7694	0.8277	0.6253	138.7155	19.3674	27.4265
	Fig. 8（e）	0.7675	0.8360	0.6227	138.7155	21.2740	22.0212
	Fig. 8（f）	0.7816	0.8835	0.6415	139.0863	22.1570	19.8925
	Fig. 8（g）	0.7937	0.9091	0.6580	102.8397	22.7167	18.6511
	Change	↑0.0243	↑0.0814	↑0.0327	↓35.8758	↑3.3493	↓8.7754

Table 1. Subjective evaluation table of cultural relic segmentation experts

View table

Table 1. Subjective evaluation table of cultural relic segmentation experts

Grade	Segmentation effect	Score
L1	Best segmentation effect	7
L2	Segmentation effect above average level	6
L3	Slightly above average segmentation effect	5
L4	Average level segmentation effect	4
L5	Segmentation effect slightly below average level	3
L6	Segmentation effect below average level	2
L7	Worst segmentation effect	1

Table 3. Subjective evaluation results

View table

Table 3. Subjective evaluation results

Group	Fig. 9（c）	Fig. 9（d）	Fig. 9（e）	Fig. 9（f）	Fig. 9（g）
Group 1	1.47	3.57	4.57	5.57	6.57
Group 2	1.60	3.60	4.60	5.60	6.60
Group 3	1.73	3.57	4.57	5.57	6.57
Group 4	3.20	3.40	4.57	5.57	6.57
Group 5	3.17	3.37	4.60	5.60	6.60

Table 4. Objective evaluation results of different algorithms on different images

View table

Table 4. Objective evaluation results of different algorithms on different images

Group	Evaluation index	Dice	Accuracy	Jaccard	HD	PSNR	RMSE
Group 1	U²-Net	0.5179	0.5013	0.3495	238.0000	18.4814	30.3717
	U-Net	0.3980	0.2583	0.2484	74.5270	18.2203	31.2986
	Robert	0.1491	0.1335	0.0805	500.2730	14.8955	45.8948
	EC	0.1786	0.2536	0.0981	502.2450	17.1941	35.2235
	Proposed algorithm	0.8975	0.9513	0.8141	34.7131	26.9814	11.4148
Group 2	U²-Net	0.7472	0.6213	0.5965	165.7227	19.8023	26.0871
	U-Net	0.4314	0.3865	0.2750	362.5960	14.8211	46.2895
	Robert	0.2409	0.1715	0.1370	175.1490	14.9190	45.7707
	EC	0.3415	0.3115	0.2059	175.8550	16.3180	38.9617
	Proposed algorithm	0.8929	0.9432	0.8064	37.8550	23.7394	16.5794
Group 3	U²-Net	0.6916	0.5324	0.5286	105.2616	20.9815	22.7754
	U-Net	0.5270	0.4199	0.3578	160.6390	17.7588	33.0067
	Robert	0.2867	0.2440	0.1673	280.6080	14.6617	47.1468
	EC	0.4052	0.5613	0.2541	281.8530	16.9127	36.3834
	Proposed algorithm	0.8890	0.9555	0.8001	69.1158	26.3602	12.2612
Group 4	U²-Net	0.8705	0.9369	0.7708	24.4131	23.9943	16.0999
	U-Net	0.2931	0.2598	0.1717	397.1860	18.9427	28.8009
	Robert	0.1651	0.1427	0.0900	30.2655	17.5668	33.7439
	EC	0.4909	0.6289	0.3253	388.8880	20.9830	22.7726
	Proposed algorithm	0.8743	0.9545	0.7767	8.0623	30.3451	7.7497
Group 5	U²-Net	0.7694	0.8277	0.6253	138.7155	19.3674	27.4265
	U-Net	0.3073	0.2664	0.1815	141.0140	16.4753	38.2624
	Robert	0.2586	0.1745	0.1485	141.0320	15.7724	41.4874
	EC	0.4738	0.3730	0.3105	140.0320	17.2989	34.8013
	Proposed algorithm	0.7937	0.9091	0.6580	102.8397	22.7167	18.6511

Table 5. Comparison of running time of different segmentation detection algorithms
View table
Table 5. Comparison of running time of different segmentation detection algorithms
Algorithm EC U-Net Robert U²-Net
Running time /ms 19 103 59 33
Algorithm U²-Net+PSA U²-Net+PSA+SA U²-Net+PSA+SA+Double layer dislocation link Proposed algorithm
Running time /ms 38 45 49 52

Tools

Get Citation

Copy Citation Text

Jiao Wang, Meng Wu, Jiankai Xiang. X-Ray Rapid Decentralized Detection Model of U²-Net with Dilated Convolution Optimization[J]. Laser & Optoelectronics Progress, 2024, 61(15): 1512006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites