A likelihood ratio test (LRT) -based modulation classifier is sensitive to unknown parameters, such as carrier frequency offset (CFO), symbol rate, etc. To deal with the limited knowledge of CFO, in this paper, a quasi-hybrid likelihood ratio test (qHLRT) -based approach is proposed for linear modulation classification. In the qHLRT algorithm, a non-maximum likelihood (ML) estimator is used to reduce the computational burden of multivariate maximization. Several of blind, non-ML CFO estimators are studied and their performance are compared with both single and multiple receiving antennas systems. It is shown that the nonlinear least-squares (NLS) CFO estimator is the best choice for the qHLRT algorithm, particularly with antenna arrays, which are introduced to combat the effect of channel fading on modulation classification.