Studies concerning the physiological significance of Ca²⁺ sparks often depend on the detection and measurement of large populations of events in noisy microscopy images. Automated detection methods have been developed to quickly and objectively distinguish potential sparks from noise artifacts. However, previously described algorithms are not suited to the reliable detection of sparks in images where the local baseline fluorescence and noise properties can vary significantly, and risk introducing additional bias when applied to such data sets. Here, we describe a new, conceptually straightforward approach to spark detection in linescans that addresses this issue by combining variance stabilization with local baseline subtraction. We also show that in addition to greatly increasing the range of images in which sparks can be automatically detected, the use of a more accurate noise model enables our algorithm to achieve similar detection sensitivities with fewer false positives than previous approaches when applied both to synthetic and experimental data sets. We propose, therefore, that it might be a useful tool for improving the reliability and objectivity of spark analysis in general, and describe how it might be further optimized for specific applications.