The image quality is a crucial factor for calculating flame temperature field based on the color CCD method. However, much unknown noise in flame images would prevent from obtaining the temperature with accuracy. To eliminate noise, the wavelet denoising with soft-threshold is considered to be effective. An important parameter for this procedure is threshold. For flame images containing much unknown noise, the paper presents using improved Generalized Cross Validation law to get optimal threshold (GCV- threshold). This procedure does not require prior knowledge of noise level and is proved to be asymptotically optimal. For comparing denoised effect, an artificial noisy image is processed by different thresholds. The results show the GCV threshold is superior to others. Furthermore, the method is applied into the industrial flame image which indicates the accuracy of the temperature measurement has been improved.