Infrared thermal wave non-destructive testing is a new type of technology that has developed rapidly in the past two decades and is widely used in many fields. However, owing to its vulnerability to influence from environmental factors and the particularity of its uneven working components, there is always a certain degree of noise pollution in the original thermal image of uncooled thermal imaging cameras; therefore, denoising the original thermal image is a key step in this technology. The traditional improved wavelet threshold denoising method is limited to the transformation of the adaptive decomposition scale of the threshold, such that the threshold function is smooth and continuous. There is no targeted method for noise variance estimation, which is the key variable of the threshold that determines the effect of wavelet threshold denoising. This study establishes a mixed noise model based on the noise characteristics of infrared images, estimates the noise variance, improves the threshold and threshold function based on the noise model, obtains the best function parameters through software, and finally analyzes the simulation results, process, and evaluation of real images. The results show that the improved wavelet threshold denoising method has a better denoising effect than the traditional threshold denoising method and partial filter denoising method.