부산대학교 도서관

In southwest China, poisoning incidents caused by accidental ingestion of wild mushroom occur frequently, so wild mushroom classification is extremely important, but the traditional CNN network lacks global modeling capabilities, and the classification effect is not as good as Transformer. Therefore, we adopted the Swin Transformer network as the backbone network of the classification. In addition, the existing wild mushroom classification tasks are greatly disturbed by the useless background in the image, which will affect its classification effect. Therefore, this paper refers to the weakly supervised object positioning module to obtain strongly differentiated regions, which including feature pyramid network, the weakly supervised object positioning selector and combiner. This module can output pixel-level feature maps, find the features of the strongly differentiated regions and fuse filtered features to enhance the effect of fine-grained visual classification. Based on the common wild mushroom in China, we constructed a dataset of wild mushrooms and conducted experiments, and after adding the weakly supervised object positioning module to the Swin Transformer network, the performance is superior and the precision is improved to 98.77%.