The lack of uniform standards for wood strength prediction in forestry is a major cause of inefficiency in Japan’s timber supply system. Three methods are proposed for wood strength prediction: the machine grading method, which requires professional instruments, equipment, and strict operational environment conditions; machine learning methods, where the quality of the wood features greatly influences the prediction of wood strength; and deep learning models can predict wood strength directly from log section images, achieving accuracy comparable to machine learning models that based feature extraction. This paper focuses on a deep learning model and designs a filtered swin-transformer for predicting wood strength from log section images. Experiments show that the accuracy of the proposed method is up to around 80%.