Domain adaptation networks for noisy image classification

In this thesis, we propose a novel semi-supervised clean-noisy datasets adaptation algorithm. We transfer the knowledge learned on clean images to unlabeled noise-distorted ones. This modification on standard deep networks produce stable classification performance on all distortion levels, which brings benefit to real-world cases. Specifically, we propose a strategy to jointly learn a shared feature encoder on the network, i.e., i) discrimination capability of network is learned by supervised training on labeled source (clean) dataset, ii) knowledge transferring is achieved by unsupervised domain adaptation to map features extracted from both domains (clean and noisy) to a common space. Our proposed network is optimized by a two-step backpropagation strategy, similar to that of Generative Adversarial Networks (GANs). We evaluate our proposed network on two popular datasets, where both show clear improvement of classification performance compared to preprocessing noisy images using the state-of-the-art denoising algorithm BM3D (up to ∼19% in average accuracy over all noise levels). Interestingly, we also observe that the proposed approach efficiently improves the feature transferability on very deep architectures, which is challenging for previous domain adaptation methods. In the future, we can also explore more challenging domain adversarial tasks like distorted image segmentation with the proposed algorithm. ii Domain adaptation networks for noisy image classification