Image classification method on class imbalance datasets using multi-scale CNN and two-stage transfer learning

Image classification tasks widely exist in many actual scenarios, including medicine, security, manufacture and finance. A major problem that hurts algorithm performance of image classification is the class imbalance of training datasets, which is caused by the difficulty in collecting minority class samples. Current methods handle this class imbalance problem from three aspects: data resample, cost-sensitive loss function and ensemble learning. However, the average accuracy of these common methods is about 95% and performance gets degenerating dramatically when the training datasets are extremely imbalanced. We propose an image classification method on class imbalance datasets using multi-scale convolutional neural network and two-stage transfer learning. Proposed methods extract multi-scale image features using convolutional kernels with different receptive fields and reuse image knowledge of other classification task to improve model representation capability using two-stage transfer strategy. Comparison experiments are carried to verify the performance of proposed methods on DAGM texture dataset, MURA medical dataset and an industrial dataset. The average accuracy obtained by proposed methods reaches about 99% which is 2.32% higher than commonly used methods over all the cases of different imbalance ratio, accuracy increase of 4.0% is achieved when some datasets are extremely imbalanced. Besides, proposed method can also achieve best accuracy of more than 99% on the industrial dataset containing only several negative samples. In addition, visualization technique is applied to prove that the accuracy boost comes from advantage of proposed architecture and training strategy.