基于迁移学习和双线性CNN的细粒度菌菇表型识别

为了对细粒度菌菇进行表型识别，在双线性卷积神经网络细粒度图像识别框架基础上，提出了一种基于迁移学习和双线性Inception-ResNet-v2网络的菌菇识别方法。利用Inception-ResNet-v2网络的特征提取能力，结合双线性汇合操作，提取菌菇图像数据的细粒度特征，采用迁移学习将ImageNet数据集上预训练的模型参数迁移到细粒度菌类表型数据集上。试验表明，在开源数据集和个人数据集上，识别精度分别达到87.15%和93.94%。开发了基于Flask框架的在线菌类表型识别系统，实现了细粒度菌菇表型的在线识别与分析。

Fine-grained Mushroom Phenotype Recognition Based on Transfer Learning and Bilinear CNN

As one of the important fungi, mushrooms have a wide variety. There are about 100000 species of fungi that have been found so far, and the phenotypes of most fungi have little difference. The identification and classification for the variety of fungi is a challenging task, which needs professional fungus expert knowledge to complete. As an edible mushroom, the study of its classification is of great importance. In order to be able to perform fine-grained phenotype recognition of mushrooms, a fine-grained mushroom recognition method was proposed based on transfer learning and bilinear convolutional neural network of Inception-ResNet-v2. For extracting the fine-grained features of mushroom image data, the Inception-ResNet-v2 network combined with bilinear convergence operation was employed. In addition, for improving the training performance, the pre-trained model parameters based on the ImageNet dataset were transferred for the fine-grained mushroom phenotype dataset using transfer learning skills. In order to evaluate the performance of the approach, extensive experiments were conducted, and the experimental results showed that the identification accuracy was 87.15% and 93.94% on the open source data set and the private data, respectively. Finally, a Flask-based online mushroom phenotype identification system was developed to facilitate the online identification and analysis of fine-grained mushroom phenotypes as well.