Can an Image Be Its Own Reference? FreeCrack for Lightweight Crack Segmentation

Thin cracks often resemble material joints, shadows, and surface texture, while foreground and background appearance shifts across images. Existing lightweight crack decoders aggregate multi-scale context but do not explicitly compare foreground and background evidence from the current input, leaving weak cracks fragmented and background textures falsely activated. We introduce FreeCrack, a compact decoder that treats each input as its own reference through an ordered evidence chain. Phase-preserving Lightweight Evidence Projection (PLEP) aligns coarse evidence to an unwarped shallow grid. In that shared geometry, Intra-image Task Prototype Mining (ITPM) constructs current-image crack and background prototypes and converts their similarity margin into dense evidence. Reliability-penalized Evidence Fusion (RPEF) then discounts dispersed token interactions before applying gated residual correction. The chain requires no support images, annotations beyond segmentation masks, or auxiliary supervision. Against the matched reproduced MixerCSeg baseline, FreeCrack raises four-benchmark average mIoU from 83.63% to 84.42% while reducing parameters by 42.5%. On TUT, mIoU rises from 81.89% to 82.68%. Fixed-width and targeted controls distinguish the effects of spatial alignment, current-image prototype source, residual correction, and reliability penalization. On DeepCrack, false positives and false negatives decrease by 11.1% and 4.8%, respectively. The decoder also improves RHACrackNet on DeepCrack and CrackMap and achieves higher Dice and IoU than matched SRF on DRIVE, spanning host architectures and a non-crack thin-structure domain. Together, these results support same-image foreground-background comparison as a compact decoder principle for resource-constrained thin-structure segmentation.