We　introduce　a　novel　concept　termed　gradient　balance　prior　to　cross-point　patterns　and　prove　three　related　theorems.　Using　this　prior,　we　develop　vanishing　power　transform　to　convert　gray-scale　images　into　response　maps,　highlighting　every　cross-point　satisfying　centrosymmetry　condition　as　humps.　By　calculating　the　hump　centers,　we　can　localize　cross-points　with　subpixel　accuracy.　Our　method　surpasses　existing　benchmarks　in　both　detection　efficiency　and　subpixel　accuracy,　as　demonstrated　in　both　simulated　and　real-world　experiments.