Gas permeability, which is measured mainly through gas permeability experiments, is a critical technical index in many engineering fields. In this study, permeability is firstly calculated based on information from a digital image and an improved permeability prediction model. The calculated results are experimentally verified. Subsequently, a self-developed image-processing program is used to extract feature parameters from a scanning electron microscopy image. Meanwhile, an extreme learning machine algorithm is used to input the image feature parameters obtained using the image-processing program into the extreme learning machine algorithm for machine learning. Additionally, we compare several typically used machine learning algorithms, which confirmed the reliability and accuracy of our algorithm. The best activation function can be obtained by comparing the predicted permeability using an appropriate number of neuron nodes. Experimental results show that the program can accurately identify the features of the microscopy image. Combining the program with an extreme learning machine neural network algorithmgas permeability results to be obtained with high accuracy. This method yields good predictions of permeability in certain cases and has been adapted to other geomaterials.