This paper proposes a method to detect infrared land mine signatures embedded in rotationally invariant colored noise. A common problem in statistical image processing is high dimensionality. This causes a need for large sets of training data. To overcome this, an alternative formulation of the generalized likelihood ratio test (GLRT) is presented. This formulation makes it possible to utilize the circular-symmetry, rendering a substantial decrease in model dimensionality and consequently, in the amount of training data needed. Simulations indicate that a significant gain in performance can be achieved compared to both the non-parameterized detector and the matched filter.