Based on the wavelet transform and its multiscale decompositions, it is possible to grasp the slight changes of characteristic information carried by image's edge on every scale. The laws have been analysed, which are followed by local maximal of wavelet transform modulus of white noise and continuous signal on different scales. An algorithm is developed to remove white noises from signals by analyzing the evolution of local maximals of wavelet transform modulus at acrossing scales. The numerical simulation proves that the algorithm is efficient. It is possible to provide an efficient approach to low-SNR radar signal detection in the future.