Prompt-Free Metal Crack Video Segmentation Using an Adapted SAM2

Given sequential metal-surface images, crack video segmentation seeks accurate crack masks and stable crack-tip localization across frames. Thin cracks with low contrasts are difficult to distinguish from the background, and specimen motion further complicates consistent segmentation across frames. To address these challenges, we propose a prompt-free Segment Anything Model 2 (SAM2) framework integrating spatial crack segmentation with a temporal consistency-assisted state update (TCSU) mechanism. The proposed method uses sparse high-level low-rank adaptation and encoder stage-end residual adapters to adapt SAM2 to thin crack structures, with boundary consistency loss strengthening boundary-sensitive training. TCSU verifies spatial evidence against registered appearance change and the previous state and recursively maintains the sequence state, including the crack mask and tip position. On SteelCrack, the proposed method achieved 78.41% miIoU and 87.02% miDice. Online experiments demonstrate the effectiveness of the proposed method in monitoring crack initiation and propagation during fatigue testing.