Texture classification is a classic problem in pattern recognition. It is an effective strategy for improving texture classification to find the texture features with both powerful discrimination and various invariant properties. In this paper, we provide a new insight into texture images, that is, texture images can be treated as quasi-periodic signals. Some new concepts such as Dominant Period Component (DPC), periodic degree (PD), and Main Frequency (MF) are proposed to characterize the properties of quasi-periodic signals. DPC controls the oscillation rate of a quasi-periodic signal and plays a key role in controlling the behavior of the whole signal. So it can serve as a key feature for texture classification. Based on this idea, we propose a new method to extract texture features. The proposed features have both powerful classification ability and rotation-illumination-invariance as well as robustness to noise. Experimental results on three texture data sets demonstrate the validity of this method. INDEX TERMS Quasi-periodic signal, main-frequency, texture feature, Hilbert marginal spectrum.