Author Affiliations
1School of Mathematics and Computer Science, Northwest Minzu University, Lanzhou, Gansu 730030, China2Key Laboratory of China's Ethnic Languages and Information Technology of Ministry of Education, Northwest Minzu University, Lanzhou, Gansu 730030, Chinashow less
Fig. 1. Line drawing of inpainting Thangka image and threshold effect images. (a) Scratch damaged of Thangka image; (b) massive damaged of Thangka image; (c) inpainting image based on sample block-based model; (d) inpainting image based on TV model ; (e)(f)line drawing of DOG operator; (g)(h), (i)(j), (k)(l) threshold effect image, φ=1.2, 1.6, and 2.0,respectively
Fig. 2. Effect of symmetric fusion image
Fig. 3. Linear diagram between color entropy and different inpainting image. (a) Color entropy diagram of massive damage; (b) color entropy diagram of scratch damage
Fig. 4. Structure of the BP-AdaBoost neural network
Fig. 5. Image restoration of damaged Thangka. (a) Scratch damaged of Thangka image; (b) massive damaged of Thangkaimage; (c) repaired image based on TV model; (d) repaired image based on sample block-based model
Fig. 6. Scatter plots of different evaluation algorithms and corresponding DMOS values on Thangka database. (a) Method in Ref. [15]; (b) method in Ref. [26]; (c) method in Ref. [16]; (d) proposed method
Fig. 7. Comparison of robustness among different algorithms
Damaged model | Inpainting model | Image number | Number of subjective evaluation |
---|
Scratch | TV model | 200 | 6 | Massive | Sample block-based model | 300 | 7 | Scratch | Sample block-based model | 300 | 8 | Massive | TV model | 200 | 9 |
|
Table 1. Composition of Thangka database
Method | Scratch-TVmodel | Massive-sampleblock-based model | Scratch-sampleblock-based model | Massive-TVmodel | All |
---|
PSNR | 0.8646 | 0.8831 | 0.9210 | 0.7515 | 0.8636 | SSIM | 0.9389 | 0.9446 | 0.9235 | 0.9046 | 0.9129 | Method in Ref. [12] | 0.8431 | 0.9391 | 0.9373 | 0.9542 | 0.9545 | Method in Ref. [13] | 0.9394 | 0.9449 | 0.9272 | 0.9246 | 0.9647 | ASVS | 0.8935 | 0.9418 | 0.9282 | 0.9424 | 0.8954 | DN | 0.9040 | 0.9291 | 0.9202 | 0.8983 | 0.9063 | Method in Ref. [15] | 0.9325 | 0.9411 | 0.9284 | 0.9428 | 0.9326 | Method in Ref. [26] | 0.9039 | 0.9287 | 0.9316 | 0.9403 | 0.9172 | Method in Ref. [16] | 0.8999 | 0.9467 | 0.9349 | 0.9435 | 0.9314 | Proposed method | 0.9327 | 0.9485 | 0.9339 | 0.9741 | 0.9463 |
|
Table 2. SROCC median of different types of repaired images in Thangka database iterate 1000 times
Method | Scratch-TVmodel | Massive-sampleblock-based model | Scratch-sampleblock-based model | Massive-TVmodel | All |
---|
PSNR | 0.8762 | 0.9029 | 0.9173 | 0.7801 | 0.8592 | SSIM | 0.9405 | 0.9363 | 0.9324 | 0.9004 | 0.9066 | Method in Ref. [12] | 0.8301 | 0.9268 | 0.9583 | 0.9640 | 0.9511 | Method in Ref. [13] | 0.9494 | 0.9449 | 0.9372 | 0.9300 | 0.9613 | ASVS | 0.8546 | 0.9356 | 0.9365 | 0.9688 | 0.8678 | DN | 0.9041 | 0.9292 | 0.9302 | 0.8983 | 0.9063 | Method in Ref. [15] | 0.9325 | 0.9211 | 0.9584 | 0.9328 | 0.9326 | Method in Ref. [26] | 0.9040 | 0.9288 | 0.9516 | 0.9403 | 0.9172 | Method in Ref. [16] | 0.8645 | 0.9367 | 0.9234 | 0.9868 | 0.9432 | Proposed method | 0.9446 | 0.9393 | 0.9555 | 0.9795 | 0.9492 |
|
Table 3. PLCC median of different types of repaired images in Thangka database iterate 1000 times
Method | Average time /s |
---|
Method in Ref. [12] | 18.342 | Method in Ref. [13] | 56.987 | ASVS | 0.0546 | DN | 4.8452 | Method in Ref. [15] | 0.0325 | Method in Ref. [26] | 63.676 | Method in Ref. [16] | 1.1645 | Proposed method | 1.2235 |
|
Table 4. Average time to process an image by different algorithms on Thangka database