论文标题
使用深层卷积神经网络对中国发现有毒和可食用的蘑菇进行分类
Using deep convolutional neural networks to classify poisonous and edible mushrooms found in China
论文作者
论文摘要
由于它们丰富的氨基酸，多糖和许多其他使人类有益的营养素，蘑菇当之无愧地作为全球和中国的饮食中的饮食美食而受欢迎。但是，如果人们错误地吃了有毒的真菌，他们可能会遭受恶心，呕吐，精神障碍，急性贫血甚至死亡的困扰。每年在中国，大约有8000人生病，而有70人因误食蘑菇而死亡。据计算，有成千上万种蘑菇，其中只有900种可食用的蘑菇，因此没有专门的知识，错误地出现有毒蘑菇的可能性很高。大多数人认为有毒蘑菇的唯一特征是鲜艳的颜色，但是，其中某些特征与这种特征不符。为了防止人们吃这些有毒的蘑菇，我们建议使用深度学习方法来指示蘑菇是否通过分析数百种可食用和有毒的蘑菇智能手机图片来有毒。我们众包一个蘑菇图像数据集，其中包含250张有毒蘑菇和200张可食用蘑菇的图像。卷积神经网络（CNN）是一种专业的人工神经网络类型，它使用称为卷积的数学操作代替至少一个层中的一般矩阵乘法，可以通过分析大量图像来产生相对精确的结果，因此非常适合我们的研究。实验结果表明，所提出的模型具有很高的信誉，可以为选择可食用真菌提供决策基础，从而降低食用有毒蘑菇引起的发病率和死亡率。我们还开源了手收集的蘑菇图像数据集，以便同行研究人员还可以部署自己的模型来推动有毒的蘑菇识别。
Because of their abundance of amino acids, polysaccharides, and many other nutrients that benefit human beings, mushrooms are deservedly popular as dietary cuisine both worldwide and in China. However, if people eat poisonous fungi by mistake, they may suffer from nausea, vomiting, mental disorder, acute anemia, or even death. Each year in China, there are around 8000 people became sick, and 70 died as a result of eating toxic mushrooms by mistake. It is counted that there are thousands of kinds of mushrooms among which only around 900 types are edible, thus without specialized knowledge, the probability of eating toxic mushrooms by mistake is very high. Most people deem that the only characteristic of poisonous mushrooms is a bright colour, however, some kinds of them do not correspond to this trait. In order to prevent people from eating these poisonous mushrooms, we propose to use deep learning methods to indicate whether a mushroom is toxic through analyzing hundreds of edible and toxic mushrooms smartphone pictures. We crowdsource a mushroom image dataset that contains 250 images of poisonous mushrooms and 200 images of edible mushrooms. The Convolutional Neural Network (CNN) is a specialized type of artificial neural networks that use a mathematical operation called convolution in place of general matrix multiplication in at least one of their layers, which can generate a relatively precise result by analyzing a huge amount of images, and thus is very suitable for our research. The experimental results demonstrate that the proposed model has high credibility and can provide a decision-making basis for the selection of edible fungi, so as to reduce the morbidity and mortality caused by eating poisonous mushrooms. We also open source our hand collected mushroom image dataset so that peer researchers can also deploy their own model to advance poisonous mushroom identification.
