Abstract: In the process of building a smart mine，the detection system has a low recognition rate for personnel violations，machine hazards，and environmental hazards under low light conditions in the mine. Considering that most of the images under the mine are real images with low pixels，high noise，and low illumination，they are not suitable for common reconstruction methods. Therefore，a reconstruction method based on fuzzy kernel estimation to construct training set is proposed to reconstruct low-resolution images into high-resolution images. The method mainly includes two parts. Firstly，the real fuzzy kernel is extracted from the image in the real data set DPED. The fuzzy kernel of the real image is estimated by using the retinex-based method，and the gradient prior information is introduced to further strengthen the edge of the fuzzy kernel. Then，the mine data set is constructed，and the extracted real fuzzy kernel is convoluted with the mine data set to construct the training set. Finally，the super-resolution reconstruction test of the real image under the mine is carried out. Experiments show that the real low-illumination image super-resolution reconstruction method based on fuzzy kernel estimation provides a higher resolution image for the de⁃ tection and recognition of the subsequent system. The algorithm is preliminarily applied to the foreign object detection in coal preparation plant，and the detection accuracy reaches 95.3%，which greatly reduces the false alarm rate and makes the coal mine production safer and more efficient.

Key words: low illumination image, super-resolution reconstruction, mine image, fuzzy kernel esti? mation