Efficient Smooth Tensor Train and Tensor Ring Completion for Image Classification Enhancement

This paper deals with studying the data completion problem for enhancing the image classification task under the pixel removal scenario. In some applications, it happens that a part of the pixels of a given image is lost due to several issues, such as corruption by outliers or artifacts and/or incompleteness due to imprecise data acquisition. This issue results in a completely wrong classification outcome using Deep Neural Networks (DNNs). In this paper we investigate the benefit of data completion in enhancing the classification accuracy of the DNN models to build more robust and stable DNN models. To this end, we propose an efficient tensor train and tensor ring completion algorithm as a preprocessing stage to deal with the mentioned problem. We apply our proposed approach to a variety of DNNs, such as AlexNet, GoogleNet, ResNet-50, and ResNet101. In particular, our experiments confirmed that even for some images with 85% missing pixels, our strategy can correctly classify the objects for all DNNs. The proposed methodology is also utilized to handle the adversarial attacks, including FGSM, PGD, Carlini-Wagner (CW), AutoAttack, and Jitter-attack. Here, we sample a subset of the pixels in the attacked image, reconstruct it, and use it for classification. Our simulation results also verified the effectiveness of this methodology.