Disclosed is a computerized method for detecting defects on a sample. The method includes: (i) receiving scan data corresponding to a pixel on the sample; (ii) computing a difference vector d based on the scan data and corresponding reference data; (iii) computing a parameter D dependent on t=Γd−(Glinear/∥Γs∥2)Γs, wherein ΓTΓ=K−1 with K being a covariance matrix corresponding to the pixel, s is a predetermined kernel characterizing a defect signal, and Glinear=s·(K−1 d) is a gaussian approximation of a likelihood ratio test expression for distinguishing the defect signal from noise, and wherein D substantially monotonically increases with ∥t∥; and (iv) computing a score q(g, D) indicative of whether the pixel is defective, wherein g is a parameter indicative of a value of Glinear and q(g, D) substantially monotonically increases with g and substantially monotonically decreases with D.