Abstract: Image⁃based coal and gangue separation is susceptible to dust，water mist，electromagnetic interference and light produced by equipment，whose factors change the environmental illuminance and further decrease image quality. This paper solves this problem by proposing a Retinex algorithm based on fast⁃guided filtering，which adjusts the brightness of the images by replacing the original illuminance component with a more suitable one. Firstly，this paper builds an image dataset consisting of lean coal and shale from Hancheng city and weakly caking coal and siltstone from Chenjia mountain mining area，and simulates the environmental illuminance of the actual working conditions by using four illuminances including 2 500，4 000，5 500 and 7 000 lux. Secondly，as the images are enhanced by adding a light factor of 0.1 with the interval in the scope of 0.1-0.9，this paper analyzes how the standard⁃deviation and entropy of the images change before and after the enhancing，as well as the difference of standard⁃deviation and the difference of entropy between coal and gangue images. After normalizing the difference of the standard deviation and the differ⁃ ence of entropy， the best light factor corresponding to the four illuminances is obtained by the maximum differ⁃ ence method. Using LSSVM as the classifier，trained and validated using entropy and standard deviation as the input vectors，the results show that the image recognition rate using the best light factor for image enhancing under the four illuminances is increased by 7.5%，8.0%，8.5%，2.0% respectively in the samples from Hancheng City，and 0.5%， 12.0%，17.0%，25.0% respectively in the samples from Chenjia mountain. The average time of image enhancement is 0.031 s.