Coal dust is the main cause of coal mine accidents. The classification and measurement of coal dust particles is very important for online detection of coal dust concentration. In recent years, particle image analysis and processing technology has been applied more and more widely, but the underground environment of coal mine is complex, coal dust image in the process of collection and transmission, will inevitably be affected by noise interference, on the subsequent particle detection. Therefore, it is very important to de-noising the image of coal dust particles. Non-local Means (NLM) denoising algorithm has a significant effect on image denoising, but for classical NLM, the use of exponential function as the kernel function will result in the loss of image details. In order to improve this defect, this paper USES cosine weighted gaussian kernel function to improve the traditional non-local mean algorithm, which can better retain the details of the denoised image. Experimental results show that the denoising performance of this algorithm is significantly better than the classical NLM algorithm, and it can better retain the detailed information in the coal dust image.