An Improved TrAdaBoost for Image Recognition of Unbalanced Plant Leaf Disease

Due to the impact of morbidity, region as well as subjective factors and so on, there are some differences in the number of each category in the plant leaf disease image data set. When a data set is unbalanced, the generalization ability of the model decreases. To improve classification accuracy, an improved TrAdaBoost is proposed. Firstly, the images in the data sets are processed called whitening to reduce input redundancy and decrease exposure, and then divided into training set and test set in a certain proportion. Secondly, in order to reduce the occurrence of negative transfer, c-means clustering is used to obtain the instances in the source domain with large similarity to the target domain. Thirdly, a balanced factor is introduced to balance the classification accuracy of minority class and majority class by adjusting the weights of instances in minority class according to the error rate in the target domain. Meanwhile the classifier in the traditional TrAdaBoost is replaced with support vector machine in order to further accommodate the image multi-classification task. To verify the effectiveness of the improved TrAdaBoost, experiments are carried out on different plant leaf disease image data sets. The experimental results show that, the proposed method is more effective in solving the problem of unbalanced data sets, compared with support vector machine, TrAdaBoost and other algorithms.