Blind Evaluation of DCT Domain Spectral Characteristics of Signal-dependent Noise in Images

The problem of noise spatial correlation is relevant for many images acquired by real life systems, and for radar images particularly. Taking into account the characteristics of such noise spectrum can contribute to a significant increase in the efficiency of further processing of these images, for example, in such operations as filtering and compression. In this paper, we propose a blind method for estimation of spectrum of spatially correlated noise in images. The method operates in the DCT domain and is applicable for different types of signal-dependent and mixed noise. Testing carried out for dual-polarization Sentinel-1 radar images has confirmed rather high accuracy of the proposed method.