By segmenting the sand sample image into single particles and identifying its components, the accuracy and efficiency of sand sample lithology analysis can be significantly improved. The existing methods for sand sample image segmentation mainly rely on traditional watershed algorithms and convolutional neural networks, but due to insufficient extraction of details from single particle rock debris contours, the mis-segmentation rate is high. Therefore, this paper proposes a single particle image segmentation and extraction method that combines convolutional neural networks and watershed algorithms, using image fusion algorithms as a bridge. Firstly, an improved Mask R-CNN network is used to quickly segment the original sand sample image and obtain its initial segmented image; Then, the initial segmented image is fused with the original sand sample image, and an improved watershed algorithm is used to segment the fusion results; Finally, using the coordinate point matching method of the original sand sample image, the resulting image obtained from watershed segmentation is corrected to complete the extraction of single particle rock debris images. The experimental results show that the accuracy of the single particle automatic segmentation and extraction method proposed in this paper is as high as 96.77%, and the model is lightweight and precise, providing a feasible and effective method for rock debris image segmentation, which can meet the needs of effectively calculating structural changes in oil reservoirs, searching for potential sediment sources, and dynamic changes of reservoirs.