Jingxian Wang’s research while affiliated with Hefei Institutes of Physical Science and other places

For the complex image segmentation problem and high complexity of model caused by digital processing technology, we first use data enhancement technology to expand dataset size, and then use deep convolutional neural networks (CNNs) multi-model integration method combined transfer learning to identify crop disease and weed. On the one hand, we make full use of the prior knowledge learned from big dataset of four single deep CNNs (VGG, Inception-v3, ResNet and DenseNet). By parameter fine-tuning, the CNNs are reused in the agricultural field to alleviate the over-fitting problem caused by insufficient data sources. On the other hand, two or more CNNs are combined by the direct average method to complete multi-model integration. We directly average the category confidence generated by different models to obtain the final prediction result. The experimental results show that the combination of deep CNNs and transfer learning is effective and the CNNs multi-model integration method can further improve the identification accuracy compared to the single CNN model. The validation accuracy of crop disease and weed dataset can reach 97.14% and 99.22% respectively by using multi-model integration and transfer learning.

Fast1 and robust recognition of crop diseases is the basis for crop disease prevention and control. It is also an important guarantee for crop yield and quality. Most crop disease recognition methods focus on improving the recognition accuracy on public datasets, but ignoring the anti-interference ability of the methods, which result in poor recognition accuracy when the real scene is applied. In this paper, we propose a high-order residual convolutional neural network (HOResNet) for accurate and robust recognizing crop diseases. Our HOResNet is capable of exploiting low-level features with object details and high-level features with abstract representation simultaneously in order to improve the anti-interference ability. Furthermore, in order to better verify the anti-interference ability of our approach, we introduce a new dataset, which contains 9,214 images of six diseases of Rice and Cucumber. This dataset is collected in the natural environment. The images in the dataset have different sizes, shooting angles, poses, backgrounds and illuminations. Extensive experimental results demonstrate that our approach achieves the highest accuracy on the datasets tested. In addition, when the input images are added to different levels of noise interference, our approach still obtains higher recognition accuracy than other methods.

As the latest breakthrough in the field of computer vision, deep convolutional neural network(CNN) is very promising for the classification of crop diseases. However, the common limitation applying the algorithm is reliance on a large amount of training data. In some cases, obtaining and labeling a large dataset might be difficult. We solve this problem both from the network size and the training mechanism. In this paper, using 2430 images from the natural environment, which contain 2 crop species and 8 diseases, 6 kinds of CNN with different depths are trained to investigate appropriate structure. In order to address the over-fitting problem caused by our small-scale dataset, we systemically analyze the performances of training from scratch and using transfer learning. In case of transfer learning, we first train PlantVillage dataset to get a pre-trained model, and then retrain our dataset based on this model to adjust parameters. The CNN with 5 convolutional layers achieves an accuracy of 90.84% by using transfer learning. Experimental results demonstrate that the combination of CNN and transfer learning is effective for crop disease images classification with small-scale dataset.