深度神经网络的研究及其在植物叶片图像识别中的应用

发布时间：2018-12-30 19:23

【摘要】：植物在我们生活中无处不在,它对整个地球生态系统的保护有非常重要的意义,而研究植物非常重要的一项内容就是精确分类。传统的识别方法需要人工对图片进行识别和标注,主观性比较强,工作量极大,进而会导致标注的准确率大大降低。因此,如何从植物叶片中自动提取图像特征并进行准确的分类已成为植物分类研究所需解决的问题。近年来,深度学习的研究表明,深层神经网络可以直接从原始图片中提取抽象特征。而且,像卷积神经网络本身还具备旋转和平移不变性,而这一特性在图像识别中显得尤其重要。因此,本文重点研究如何将卷积神经网络应用于植物叶片图像识别上,并构建一个全卷积神经网络模型。针对以往植物叶片识别方法识别率偏低的问题,本文提出的具体研究内容如下:(1)设计了一种全卷积深度神经网络的图像特征提取方法。首先对图片进行预处理,包括图像切割、增强、旋转和透视等来减少一些计算量。然后,使用卷积神经网络的算法自动提取图像特征。最后,用多层感知机来替换传统神经网络的线性滤波器,由多层感知机组成的非线性卷积层把输入的局部数据通过非线性的激活函数进行映射,全连接的网络非常容易过拟合,而非线性的卷积由于权值共享,本身就具备一定防止过拟合的能力,这样得到的卷积神经网络能有效地从非线性可分的数据中抽象出所需的特征。(2)提出一种使用PRe LU代替传统ReLU激活函数的深度神经网络方法。ReLU(线性纠正函数)加快了训练过程的衔接,而且比传统的S形能更好地加速收敛。但在植物叶片图像训练中会出现梯度弥散情况,所以本文使用PReLU的激活函数,它能够自适应学习整流器的参数,并且提高准确率而可以忽略额外的计算成本。(3)通过实验及实验结果最终验证了深度卷积神经网络在植物叶片图像识别问题上是可行的。并且与其它一些神经网络算法相比较,本文改进的深度卷积神经网络在植物叶片图像识别上的识别率更高。
[Abstract]:Plants are everywhere in our lives. They are very important to the protection of the whole earth ecosystem, and one of the most important things to study plants is to classify them accurately. The traditional recognition method needs to recognize and annotate the picture manually. The subjectivity is strong and the workload is great, which will lead to the reduction of the accuracy of the tagging. Therefore, how to automatically extract image features from plant leaves and classify them accurately has become a problem to be solved in plant classification research. In recent years, the deep learning research shows that the deep neural network can extract abstract features directly from the original images. Moreover, the image convolution neural network itself has rotation and translation invariance, which is especially important in image recognition. Therefore, this paper focuses on how to apply convolution neural network to plant leaf image recognition and construct a full convolution neural network model. Aiming at the problem of low recognition rate of plant leaf recognition methods in the past, the specific research contents are as follows: (1) an image feature extraction method based on full convolution depth neural network is designed. First, the image is preprocessed, including image cutting, enhancement, rotation and perspective. Then, the algorithm of convolution neural network is used to extract image features automatically. Finally, the linear filter of traditional neural network is replaced by multilayer perceptron, and the input local data is mapped by nonlinear activation function in the nonlinear convolution layer composed of multilayer perceptron, so the fully connected network can be easily overfitted. Because of the sharing of weights, nonlinear convolution has the ability to prevent over-fitting. Thus the convolutional neural network can effectively extract the required characteristics from the nonlinear separable data. (2) A depth neural network method. ReLU (linear correction using PRe LU instead of the traditional ReLU activation function is proposed. Function) speeds up the convergence of the training process, Moreover, it can accelerate the convergence better than the traditional S form. However, gradient dispersion will occur in the image training of plant leaves, so we use the activation function of PReLU, which can learn the parameters of rectifier adaptively. And the accuracy can be improved and the extra calculation cost can be neglected. (3) the experiment and experimental results show that the deep convolution neural network is feasible in plant leaf image recognition. Compared with other neural network algorithms, the improved deep convolution neural network has a higher recognition rate in plant leaf image recognition.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP183

【相似文献】