Abstract: 9Ni steel serves as a critical material for liquefied natural gas （LNG）storage and transportation equipment. The accurate determination of its fracture modes heavily relies on high-quality scanning electron microscope （SEM） fracture images. However, image noise introduced during testing can obscure key microscopic morphological features. This not only significantly increases the difficulty and error rate of manual identification but also reduces the effectiveness of traditional image processing methods. Therefore, developing an adaptive denoising algorithm for fracture images that can effectively handle high noise interference is key prerequisite for providing high-quality data foundation for intelligent recognition. To address high-noise fracture images, an improved recursive kernel mean distance（RKMD） adaptive denoising algorithm was proposed. By incorporating wavelet transform through a local adaptive strategy, this algorithm effectively mitigates the excessive smoothing of critical features such as edges and textures during the denoising process of fracture images. By constructing an objective function aimed at minimizing residual kurtosis and integrating multi-scale edge-aware mechanism, the threshold selection strategy was dynamically optimized. This achieves a better balance between noise suppression and the preservation of edge topology in fracture features. Based on a dataset constructed from 3 050 sets of high-resolution fracture images, data-driven intelligent discrimination of fracture modes was achieved. Experimental results show that under extreme conditions with noise standard deviation as high as 30%, the proposed algorithm achieves a peak signal-to-noise ratio of 26.69 dB and a structural similarity index of 0.768 2, demonstrating significantly improves performance compared to traditional wavelet threshold denoising methods. The proposed denoising algorithm provides a clear and reliable image data foundation for subsequent intelligent recognition of fracture morphology. It notably enhances the robustness and accuracy of the fracture performance analysis process for 9Ni steel and also supplies high-quality standardized training data for subsequent deep learning models.