Research On Plant Disease Recognition Technology Based On Rectified Meta-Learning And Zero-shot Learning

Plant diseases serve as one of main threats to food security and crop production.It is thus valuable to exploit recent advances of artificial intelligence to assist plant disease diagnosis.One popular approach is to transform this problem as a leaf image classification task,which can be then addressed by the the powerful implicit feature extraction ability of convolutional neural networks(CNNs).However,the perfor-mance of CNN-based classification approach depends on a large amount of high-quality manually labeled training data,which are inevitably introduced noise on labels in practice,leading to model overfitting and performance degradation.In addition,due to the different rarity of plant species and diseases,there will be extreme unbalanced(zero-shot learning)problems with very few or no pictures of certain species and diseases in the process of obtaining the data set.In response to these two problems,we have proposed the following two solutions:1)For the noisy label,this paper using the self-learning ability of meta-learning and rectification module,a self-learning framework based on modified meta-learning is proposed to improve the robustness of the network.our method contains two phases the first phase is to train a CNN-based model to predict the classification results according to original noisy labels,and the second phase to combine meta-learning with rectification module to improve network tolerance of noise.Specifically,first generate multiple sets of synthetic mini-batch with pseudo-labels,then force the meta-learning to update the prediction of the network to be consistent with the prediction of the first phase network,and finally use a rectification module similar to the attention mechanism to give more penalty on highly-biased samples,while less penalty on unbiased ones,leading to less proneness to overfitting.Furthermore,our method is free on assumption of label noise distribution,and can be used as a plug-and-play module,which can be embedded into any deep models optimized by gradient descent based method.The experimental results on the PlantVillage dataset are conducted to demonstrate the superior performance of our algorithm over the state-of-the-arts.2)For the zero-shot learning,this paper combines user-defined attributes and latent attributes of visual features,a zero-shot learning framework based on discriminative learning is proposed,which can identify unseen classes of plants.In zero-shot learning,to solve the problem that user-defined attributes cannot accurately distinguish fine-grained samples,we propose a latent attribute based on visual features to assist user-defined attributes and generate a more comprehensive and accurate attribute table.The framework first uses convolutional neural networks(CNNs)to extract image features,and then embeds them into the enhanced embedding space that introduces latent attributes.Finally,it combines two attributes to jointly predict unseen classes.Extensive experimental results show that the proposed latent attributes have stronger discriminatory properties than the user-defined attributes,and the proposed algorithm has better performance than the existing zero-shot learning algorithm.The above two solutions have improved technologies such as meta-learning and zero-shot learning,and have added innovative technologies such as rectification module and latent attributes.Experiments show that our method can solve the zero-shot learning and noisy label in plant disease classification.