An inspection method of timber surface knots was developed based on Faster R-CNN model. The new method made the selection of knot features data-driven, adapted variation of wood species and knot types, and improved inspection accuracy efficiently. The test results showed that the positions and sizes of the knots were detected with 94.0% of inspection accuracy. The images of the four typical samples showed that the maximum errors of knot positions and sizes were 5 pixels and 7 pixels, respectively. The Faster R-CNN method was superior to OSTU-based method in the study.