Machine-made sand is the fine aggregate for machine-made sand concrete. The quality of machine-made sand, determined by the stone powder content, has a significant impact on the strength, workability, durability, and other performance of machine-made sand concrete. Considering that with low accuracy and long duration, the traditional stone powder detection methods are cumbersome and difficult to quantify, this study proposes an improved UNet model based on the characteristics of machine-made sand. First, optical microscope equipment is used to collect images of machine-made sand particles, and these images are preprocessed by means of contrast enhancement, the look-up table algorithm, low-pass filtering, etc. Then, the deep residual and attention mechanism module is introduced to build an improved UNet model. Finally, segmentation and quantitative calculation are conducted on the stone powder in machine-made sand. The results show that the segmentation accuracy of the deep neural network constructed in this paper on the machine-made sand training dataset and the verification dataset is as high as 95.2% and 95.94%, respectively, and compared to the UNet, FCN, and Res-UNet methods, this method has significantly improved the segmentation effect on the same dataset.