A Dynamic Feature Weighting Method for Mangrove Pests Image Classification with Heavy-tailed Distributions
Zixu Wang, Weiguo Gong, Weihong Li · 2020
Most of the data in agriculture and forest obey heavy-tailed distribution, in which the data distribution is extremely imbalanced and the amount of most classes is poor. In this paper, we propose a feature centroid based dynamic feature weighting method between head and tail classes for mangrove pest images classification. In the proposed method, firstly, we design a convolutional neural network to extract the clustering centroid as a representation of the subclass. Secondly, we use the centroid feature memory to achieve the dynamic feature weighting through extraction, injection and fusion, which enrich the feature space of the tail classes while retaining identity. Thirdly, to localize the pests in natural background, we optimize the weakly-supervised localization algorithm with a multiscale local maximum index strategy to extract saliency maps. Furthermore, to carry out our research we establish a heavy-tailed distributed image dataset named MIPDGC based on mangrove forest pest. The experimental results prove that the dynamic feature weighting method can obtain higher classification accuracy and better performance than other general methods on the classification accuracy of the MIPDGC.