【摘要】 为分析果树在种植过程中病害的程度和种类,本文提出一种基于改进深度残差网络的果树叶片病害图像识别方法。该网络模型在传统残差神经网络的基础上,通过多尺寸的卷积核代替骨干网络中的7×7卷积核,既增加了网络的宽度,也增加了网络对尺度的适应性。带泄露修正线性单元(Leaky ReLU)激活函数用于替换修正线性单元(Recitified Linear Unit, ReLU)激活函数,该函数以ReLU函数为基础,在函数的负半轴上引入一个非零斜率(Leaky),解决了ReLU函数引起的神经元死亡现象。在平均池化层和全连接层之间加入Dropout(按照一定的概率将神经网络单元暂时从网络中丢弃)操作,合理设置阈值,可以有效地防止卷积神经网络的过拟合。最后,引入SE注意力机制进一步提高网络模型的识别精度。在公共数据集Plant Village(植物村)的实验表明,改进的深度残差网络模型能够很好地识别果树叶片病害,平均准确率可达到99.4%。更多还原

【Abstract】 In order to analyze the extent and types of diseases in fruit trees during planting, this paper proposes an image recognition method of fruit tree leaf diseases based on improved deep residual network. Based on the traditional residual neural network, this network model replaces the 7×7 convolution kernel in the backbone network with a multi-size convolution kernel, which not only increases the width of the network, but also increases the adaptability of the network to scale. The Leaky ReLU activation function is used to replace the Recitified Linear Unit(ReLU) activation function. This function is based on the ReLU function and introduces a non-zero slope on the negative half axis of the function(Leaky), solved the neuron death phenomenon caused by the ReLU function. Adding Dropout(temporarily discarding the neural network unit from the network according to a certain probability) operation between the average pooling layer and the fully connected layer, and setting the threshold reasonably can effectively prevent the over-fitting of the convolutional neural network. Finally, the SE attention mechanism is introduced to further improve the recognition accuracy of the network model. Experiments in the public data set Plant Village show that the improved deep residual network model can identify fruit tree leaf diseases well, with an average accuracy rate of 99.4%.更多还原