Classification images provide compelling insight into the strategies used by observers in psychophysical tasks. However, because of the high-dimensional nature of classification images and the limited quantity of trials that can practically be performed, classification images are often too noisy to be useful unless denoising strategies are adopted. Here we propose a method of estimating classification images by the use of sparse priors in smooth bases and generalized linear models (GLMs). Sparse priors in a smooth basis are used to impose assumptions about the simplicity of observers' internal templates, and they naturally generalize commonly used methods such as smoothing and thresholding. The use of GLMs in this context provides a number of advantages over classic estimation techniques, including the possibility of using stimuli with non-Gaussian statistics, such as natural textures. Using simulations, we show that our method recovers classification images that are typically less noisy and more accurate for a smaller number of trials than previously published techniques. Finally, we have verified the efficiency and accuracy of our approach with psychophysical data from a human observer.