Aiming at the problem that the background of ground penetrating radar (GPR) image contains a lot of noise, which leads to the difficulty of effective signal extraction and recognition, a GPR denoising method based on improved Markov random field combined with Otsu algorithm is proposed.Firstly, the Otsu algorithm was used to binarize the GPR image data, and the foreground and background images were divided. Then, the improved Markov random field combined with Iterated Conditional Mode algorithm (ICM) was used to denoise the image, and the binary image containing only the effective signal was obtained. The binary image was combined with the noisy image, and the median filtering method was used to reduce the noise, finally the optimized GPR image was obtained. According to the experimental results, this method has a good effect on the ground penetrating radar data noise reduction and effective signal extraction. Compared with other traditional and original Markov random field denoising methods, it performs best in various GPR evaluation methods. Specifically, it has the highest fitting degree in the signal single-channel waveform contrast, the peak signal-to-noise ratio reaches 52.5281dB, and the structural similarity is 0.9981. Meanwhile, the image waveform recognition become clearer by using the filter image which combined the binary image. Therefore, this method has a certain practical value in the use of GPR detection and flaw detection engineering projects.