Multi-task Image Classification Method Based On Sample Generation

Vision is an important medium for human beings to understand the world.Similarly,computer vision is also an important way of the machine cognitive the world.Image classification has always been a concern in the field of computer vision.Deep learning technology provides many solutions for image classification problems.But behind the success of the deep learning model,there is a large amount of annotation data,which are manually labelled and very expensive.Data augmentation is a widely used technique to overcome this issue,which enlarges the training samples with label invariant transformations.However,the diversity of images generated by standard data augmentation is quite limited and thus the final improvement on classification accuracy is not much.Generative Adversarial Network(GAN)has shown powerful ability in image generation.The samples generated by GAN can introduce more diversity to the generated samples,which lead to more improvement on the classification result.Most of the GANs are designed for general image tasks.But for the fine-grained image classification problem,the training samples of each category is scarce and the differences between images of distinct categories are quite slight.This paper proposes a modified conditional GAN called F-CGAN.F-CGAN adds the multi-level features of the original image samples to the generator,so that the generated samples have similar details to the original samples.This reduces the number of samples required for training and ensures that the class labels of the generated samples are consistent with the original ones.Experiments show that F-CGAN can generate meaningful samples even in the fine-grained scenarios.On the other hand,after obtaining the synthetic images by F-CGAN,this paper discuss how to effectively leverage them for improving the classification accuracy.According to our experiments,directly combining them into training set barely leads to any improvements.The reason is that though F-CGAN can mimic the distribution of each class,there inevitably exist some divergences between them.Hence,this paper propose a multi-task classifier,which aims to classify the original training samples and synthetic images simultaneously.The two tasks share the hidden layers but have their own specific output layers.Empirical results show that by using the multi-task classifier,we can obtain 1%-2% improvement on classification accuracy for the fine-grained datasets.