Fidelity Estimation Improves Noisy-Image Classification With Pretrained Networks

Image classification has significantly improved using deep learning. This is mainly due to convolutional neural networks (CNNs) that are capable of learning rich feature extractors from large datasets. However, most methods trained on clean images and not robust when handling noisy ones, even if a restoration preprocessing step applied. While novel address this problem, they rely modified thus necessitate retraining. We instead propose method can be applied $pretrained$ classifier. Our exploits fidelity map estimate fused into the internal representations extractor, thereby guiding attention network making it more data. improve noisy-image (NIC) results by margins, especially at high noise levels, come close fully retrained approaches. Furthermore, as proof concept, we show our oracle outperform methods, whether or restored images.