Image segmentation which identifies the regions of image objects by thresholding gray levels of pixels on the basis of histogram information is treated. The main aim of this paper is to obtain the statistical performance of the image thresholding method whose threshold is determined on the basis of the squared-distortion criterion and Lloyd's algorithm for the normalized gray level histogram. In order to obtain the performance, it is assumed that the normalized histogram is governed by a mixture of generalized Gaussian p.d.f.'s, which provides a wide range of expression as histogram approximations of real images. The performance of the method using Lloy's algorithm is evaluated through the misclassification probabilities for variation of the parameters of the mixed p.d.f. in comparison with optimum thresholding and valley thresholding. Moreover, a discrete histogram model obtained by quantizing the continuous histogram is studied. These results offer us the global and systematic performance of the image thresholding methods.