基于轻量级神经网络的农作物病害识别算法

Recognition Algorithm for Crop Disease based on Lightweight Neural Network

应用深度学习的图像分析技术,可较早地、无损地检测作物病害,但移动端计算资源的有限性限制了深度学习在移动端的应用和发展。利用迁移学习方法,进行多种神经网络的预训练,将其在ImageNet图像数据集上学到的知识迁移运用到多种农作物数据集及番茄单作物数据集的多种病害识别上,并进行多个深度学习模型在多种作物数据集的计算复杂度、识别效果及计算速度的对比。通过对比发现:Xception模型的计算准确率比较高,计算复杂度稍复杂些;当应用场景对计算准确率的要求不是很高的情况下,ShuffleNet V20.5x模型在计算复杂程度、计算速度的综合表现相对较好,比较适合在移动端进行移植;接着,通过对ShuffleNet V20.5x采用ReLU和LeakyReLU激活函数进行训练和验证分析,发现当采用LeakyReLU激活函数替代原有的ReLU激活函数构建Shuffle Net V20.5x模型,可以改进Shuffle Net V20.5x模型,并能稍微提高识别的准确率,由85.6%提高到86.5%。最后将改进后的ShuffleNet V20.5x模型,移植到移动终端并进行测试。

The application of deep learning image analysis technology can detect crop diseases earlier and non destructively,but it was limited by the computing resources of mobile terminal.The knowledge learned from ImageNet image data set is transferred and applied to the identification of multiple diseases in multiple crop data sets and tomato single crop data sets.The computational complexity,recognition effect and calculation speed of multiple deep learning models in multiple crop data sets were compared.Through comparison,it is found that:Xception model has higher calculation accuracy and slightly more complex calculation complexity.When the application scenario does not require high calculation accuracy,ShuffleNet V20.5x model has relatively good comprehensive performance in calculation complexity and calculation speed,which is more suitable for transplantation in mobile terminal.Then,to compare the training and verification performances,the ShuffleNet V20.5x model was constructed with ReLU and LeakyReLU activation function separately.It is found that the ShuffleNet V20.5x model can be improved by using LeakyReLU activation function instead of the original ReLU activation function,and the recognition accuracy can be improved from 85.6%to 86.5%.Finally,the improved ShuffleNet V20.5x model is transplanted to the mobile terminal and tested.