Correlation-Filter Enhanced Meta-Learning for Classification of Biomedical Images

Recent deep learning methods have demonstrated remarkable impact on the classification of biomedical images. In this paper, we proposed a correlation-filter enhanced meta-learning approach for the classification of biomedical images. Firstly, in the training stage, we use the training samples to optimize the model parameters of meta-learning. Secondly, in the testing stage, we utilize the data samples of the new task to generalize the model parameters. Thirdly, the nearest neighbor image from one sample batch is searched for the new instance image, with the classifying score provided by the meta-learning model. Fourthly, the template of the circular cross-correlation filter is optimized in the Fourier domain, using the new instance image and its nearest neighbor image. Fifthly, the support weight of the sample batch is calculated for the classified label by the meta-learning model. Finally, we propose the multi-batch voting mechanism to decide the label of the new instance based on the correlation-filter template. Experiments on the classification of biomedical images demonstrated the effectiveness of our approach, compared with other state-of-the-art methods.