The invention discloses an automatic pulmonary tuberculosis detection system on an X-ray chest radiograph based on a characteristic pyramid network (FPN). An X-ray chest radiograph of pulmonary tuberculosis is marked; a Faster R-CNN network learning module with an FPN as the rear end is adopted for training and learning, pulmonary tuberculosis lesion symptoms are mastered, and the automatic diagnosis and detection capacity of the pulmonary tuberculosis lesion symptoms is obtained, so that the automatic detection, positioning and probability prediction of the pulmonary tuberculosis lesion are achieved, and a final pulmonary tuberculosis detection result is obtained. The FPN serves as the rear end of the detection network, the semantic features in a multi-scale network layer can be better combined, each layer is independently predicted, and fusion is finally carried out, so that focuses of different scales can be better detected. Based on a recognition technology of the deep learning network to the digital image, the automatic detection, positioning and probability prediction of the tuberculosis focus are realized, the accuracy of the focus detection is improved, and the risk of thedelayed treatment of a tuberculosis patient is reduced.