基于深度残差网络与迁移学习的毒蕈图像识别

我国毒蕈种类繁多且分布广泛,经常有人因无法鉴别毒蕈和可食用菌而误食毒蕈,导致身体健康甚至生命安全受到严重威胁。为了减少毒蕈中毒事件的发生,本文以中国常见毒蕈为研究对象,提出基于深度残差网络与迁移学习的毒蕈图像识别方法。首先通过互联网途径获取常见种类的毒蕈和非毒蕈的图像,经筛选后得到18种毒蕈和5种非毒蕈共14669张图像,使用数据增强扩充数据量,建立中国常见毒蕈图像数据集。然后以ResNet-152为预训练网络模型,采用基于模型的迁移学习方法,构建出毒蕈图像识别的模型结构,以Adam算法为模型优化方法,最后通过k折交叉验证进行模型训练。试验结果表明,毒蕈图像识别模型Top-1和Top-5准确率分别为92.17%和97.35%,对于常见毒蕈图像具有较高的识别率,可以有效的帮助人们避免误食毒蕈,为毒蕈识别研究提供新的方法。

Toadstool Image Recognition Based on Deep Residual Network and Transfer Learning

Various toadstools spreading widely in China are often taken by mistake because people lack knowledge of distinguishing them from edible fungi.In order to reduce the occurrence of toadstool poisoning events,this paper takes the common toadstool species in China as the research object and puts forward the toadstool image recognition method based on deep residual network and transfer learning.Firstly,the images of common species of toadstools and non-toadstools were obtained from Internet,a total of 14669 images of 18 species of toadstools and 5 species of non-toadstools were available after screening,and with data augmentation method,the data scale in China was still increasing.Thus,the image data set of common toadstools can be established.Thereafter,ResNet-152 network as the pre-training model,the model-based transfer learning method and Adam algorithm as the model optimization method were applied to construct the model structure of toadstool image recognition.Finally,the model training was carried out by the k-fold cross validation.The experimental results show that the recognition rate of common toadstool images is high.The accuracy of Top-1 and Top-5 of toadstool image recognition model even can reach 92.17%and 97.35%respectively.Therefore,this model can effectively help people avoid eating the toadstools,and enrich the studies of toadstool recognition.