CrackEvo
Crack segmentation, measurement and evolution analysis for laboratory tests on rock-like materials. 岩土类材料室内试验的裂纹识别、测量与演化分析。
Segment an image 单图分割
Upload one frame; get the crack mask, crack area, skeleton length, mean width and number of cracks, in pixels or mm.
Analyse a sequence 序列分析
Upload the frames of a test or a short video; the initial state (pen marks, pores, edges, old cracks) is subtracted and you get the growth curves of new cracks and the key frames from onset to 90 % growth, with the stress level of each when the loading is force- or stress-controlled.
Results 结果
Accuracy under a leakage-free protocol: every test specimen is held out as a whole, never split frame by frame.
Data behind CrackEvo 数据
| Laboratory tests of the CrackEvo team | tests | frames | labelled frames |
|---|---|---|---|
| Concrete cuboids, uniaxial compression (manual hydraulic), 1920×1200 camera | 4 | 109 | 109 |
| Cement mortar cubes, sand ratio 0–60 % × loading rate 0.5/0.7/0.9 MPa/s, 1920×1200 camera at 10 fps | 21 | 12,965 | 255 |
| Mortar and paste cubes with sand grading, pebble aggregate or steel balls, and rock samples; DV + phone video 38 DV + 33 phone videos; frames = DV seconds at 1 fps | 38 | 32,581 | 0 |
| SHPB Brazilian discs of coal (two bedding orientations) and sandstone, 50,000 fps | 32 | 468 | 468 |
Auxiliary data, used only to help training or as external checks: SHPB high-speed sequences at 100,000 fps and digital-image-correlation sequences from partner laboratories, and public pavement/concrete crack datasets (training only). The laboratory dataset is not yet public.
How it works 原理
A UNet with a ResNet-34 encoder segments cracks in 512-pixel windows that slide over the full image at its original resolution. The model runs in your browser with ONNX Runtime Web (WebGPU when available), so images never leave your computer. Crack evolution is summarised by the area of new cracks: pixels that are already predicted in the first frames are treated as static and removed, a crack must persist in two consecutive frames, and the curve is fitted as non-decreasing because visible cracks do not close under increasing load. Key frames are where the fitted growth reaches 5 % (onset) and 10–90 % of its value before failure.
Roadmap: image and sequence analysis (online) · leakage-free cross-validated benchmark (done) · crack evolution against applied stress (in progress) · synchronised crack–electromagnetic–microseismic analysis (planned) · physics-constrained crack-evolution prediction (CrackGPT, planned).