PSI - Issue 64

Jie Wang et al. / Procedia Structural Integrity 64 (2024) 1326–1333 Author name / Structural Integrity Procedia 00 (2019) 000–000

1329

4

Upsampling with higher resolution . The number of matched point pairs was determined in coarse matching module, which was limited by resolution of the feature map. Deconvolution layers with higher resolution were added in the feature map extraction module, extracting a feature map of 1/1 size to achieve increased number of matched point pairs. This new feature map retained the original resolution of the input image and was utilized in both the coarse and fine matching modules. The model architecture is depicted in Fig. 2.

Fig. 2. Structure of the feature matching model.

Dividing images into overlapping patches . The model could not handle input images with excessively high resolution, yet lower-resolution images significantly reduced the number of matched point pairs. Therefore, the original image was divided into multiple overlapping patches, with each pair of patches serving as input of the model. When subjected to dynamic loading, points in the edge regions of patches may move beyond the boundaries, an overlapping region was introduced between patches. Each patch overlapped with its adjacent patch in four boundaries (if exists). When the points in a patch moved outside the region, they could be matched in the next pair of patches. Displacement smoothness constraint . After matching of feature points, it was necessary to eliminate false matches. It was believed that displacement of any point should be similar to that of one or more neighboring points. Difference in displacement of each point and that of its surrounding points was calculated. When the difference exceeded a certain threshold, which was set as 0.5 pixel in this study, it was recognized as a false match and was therefore discarded. On average, more than 3% of the matching pairs were false and eliminated in this step for a more accurate result. 2.2. Evaluation of the global displacement field Since the discrete matched point pairs were irregularly distributed, a special interpolation method was used to evaluate the global displacement field based on the displacement of matched points. (1) The matched points were used as vertices to generate Delaunay triangles. Each point that needed to be interpolated lay in a Delaunay triangle. (2) Calculate the barycentric coordinates ( , , ) of the point to be interpolated in the triangle, satisfying the following equation:

A x x x x y y y y α β γ α β γ α β γ = + + = + + + + = B A B 1

C

(1)

C