Steel slag stability testing is the key to achieve safe resource utilization of steel slag. To address the issues of low efficiency and inadequate representativeness in stability testing methods, a steel slag image classification model called SE-ConvNeXt based on convolutional neural networks was proposed. The classification model adds channel attention mechanism squeeze and excitation network (SE-Net) to the ConvNeXt network for the image features of steel slag sand. Compared to the original ConvNeXt and other convolutional neural network models, SE-ConvNeXt achieves faster convergence, more stable training, and higher accuracy. The networks were trained using steel slag sand images of two grain sizes, 2.36 ~ 4.75 mm and 1.18 ~ 2.36 mm, respectively, and analyzed in conjunction with the change law of pulverization rate. The results show that the model predicts the accuracy of the steel slag sand image dataset of two particle sizes 92.5% and 94%, respectively, and the steel slag sand image changes gradually with the increase of steam aging time, and then the degree of image change tends to stabilize, similar to the law of change of pulverization rate. It can be seen that steam-treated steel slag sand can be evaluated by the volume stability of steel slag sand images.