Abstract: In the detection of log volume, the accuracy and stability of the intelligent detection system are easily affected by the factors such as end cutting marks, cracks, shadows and so on. For a long time, the end face recognition and positioning is a difficult problem. For this reason, through the principle analysis and experimental verification of multiple YOLO version models, this paper has fused the advantages of multiple YOLO version models. On the basis of the yolo3 backbone network, using data enhancement, feature fusion, loss function and other optimization methods, an end-to-end depth learning model YOLO-Raw Wood（YOLO-RW） is constructed, which is more suitable for log detection and is used for accurate recognition and localization of log volume images. In order to test the performance of the YOLO-RW model, this paper has designed several sets of data experiments. The experimental results show that the YOLO-RW model has higher end face recognition accuracy and robustness than the benchmark model, and is 6.95% and 2.38% higher than the benchmark model in terms of preparation rate and recall rate. Therefore, the YOLO-RW model has a good application value in the field of log volume detection, and can also provide a reference for the research of related target recognition fields.