Research On Image Recognition Of Crop Disease Based On Instance And Parameter Transfer

In the process of crop cultivation,it's the key to timely and effectively diagnose diseases for the healthy growth of crop.There are two ways to diagnose crop diseases at present.One is based on manual work which is a common way.The other one is based on traditional machine learning which is developing rapidly.However,it's easy to integrate into the individual subjective consciousness for the former and it usually need to satisfy the assumption that the training data which is a large scale of labeled samples and test data should be under the same distribution for the latter.In the practical application,on the one hand,the cost of image collection and marking of some diseases is high,so it's difficult to satisfy the assumption of traditional machine learning.On the other hand,there are a plenty of labeled samples which are related to crop disease image.In this situation,transfer learning is introduced to identify several diseases of cucumber and rice in the paper.(1)In the view of the problem that using traditional machine learning is not easy to obtain the ideal recognition result with a small number of crop disease images,the paper develops a method of instance-based transfer learning.In order to timely and effectively segment the lesion area from cucumber and rice leaves with simple background,the double Otsu method that chooses the components in different color space to carry out gray image conversion is employed.Then 19 parameters including color,texture and shape feature of spot image are extracted as feature vectors.Finally,the paper analyses the problem of TrAdaBoost algorithm and designs a method to optimize the auxiliary data based on K-nearest neighbor algorithm.The experiment results reveal that the proposed approach can dropout some auxiliary data with lower similarity to the target data and make use of the useful knowledge of auxiliary data,so as to improve the recognition effect of crop disease image.(2)In the view of the problem of that training convolutional neural network with small sample of crop disease image is easy to produce over-fitting,the paper proposes a recognition method of parameter-based transfer learning with convolutional neural network.Here,two strategies are adopted to pre-process 8 diseases images in cucumber and rice with complex background and inconsistent size.One strategy does not change the number of target data,and another strategy achieves the expansion of the number of target data via cropping images two times.Two networks,namely AlexNet and VGGNet,are used to train two pre-trained models with PlantVillage that is an open source dataset.Then the original networks are optimized by combining batch normalization with DisturbLabel algorithm and the pre-processed target data is applied to fine-tune the pre-trained models.The paper compares the proposed method with training network from scratch and traditional machine learning method,and results indicate that the proposed method outperforms other two methods by about 1%to 9%in terms of classification accuracy rate.